UniversitasAI™

The Vision

Why higher education needs an AI operating system — and why now.

A modern university is one of the most complex organizations on Earth. Thousands of students, hundreds of employees, dozens of departments — each with its own workflows, deadlines, regulations, and stakeholders. Admissions is running pipeline analytics. Finance is tracking receivables across multiple currencies. Student Success is trying to identify at-risk students before it’s too late. HR is managing visa renewals, leave requests, and contract cycles. The registrar is resolving scheduling conflicts across hundreds of course sections. Marketing is timing social campaigns across global time zones. Compliance is preparing for accreditation reviews. And career services is coordinating employer placements for graduating cohorts.

Each of these teams runs on different systems with different data. Between the silos, thousands of small but critical tasks fall through the cracks every semester:

These aren’t technology failures. They aren’t caused by incompetent staff. They’re coordination, timing, and complexity failures. Each department is doing its job well — but university operations are inherently cross-departmental, time-sensitive, and statistically subtle. No human team, however talented, can monitor everything, notice every pattern, and act on every opportunity in real time.

This is the premise behind UniversitasAI: an AI operating system that doesn’t just alert humans to problems, but actively solves them — while keeping humans firmly in control of the decisions that matter most. Not a chatbot. Not a dashboard. Not a reporting tool. An autonomous institutional intelligence that operates across every department, detects situations using five different analytical methods, makes decisions through a graduated safety system, executes actions through real integrations (payments, e-signatures, notifications, government reports), measures real-world outcomes, and continuously calibrates its own judgment.

The Scale of What This Solves

To understand UniversitasAI, you need to understand the operational scale of a single university:

Addressing AI Reliability

The Opportunity

Protected Innovation — Patent Pending

The core architecture of UniversitasAI is the subject of a provisional patent application filed with the United States Patent and Trademark Office (USPTO). The patent covers a novel system and method for autonomous institutional decision-making — specifically the combination of technologies that, to our knowledge, has never been assembled in a single platform:

What is UniversitasAI?

Not a chatbot. Not a dashboard. A multi-layered intelligence system with 25 specialized AI agents, 106 autonomous scanners, a real-time communication mesh, and three independent learning mechanisms.

UniversitasAI is a unified AI operating system built from four architectural layers, each sophisticated in its own right, and together forming an institutional intelligence that no single technology — not an ERP, LMS, SIS, CRM, not a BI tool, not a chatbot — can replicate.

Layer 1: The Sensing System — 106 Scanners, 5 Detection Methods

Every 10 minutes, a background process sweeps the entire institution. This isn’t a simple cron job checking a few values — it’s 106 specialized scanners organized into five fundamentally different detection methodologies:

Layer 2: The Decision Engine — Graduated Autonomy

When a scanner detects a situation, it doesn’t simply “look at a value and act.” The situation enters a multi-stage decision pipeline that evaluates risk, consults peer agents, checks policy constraints, and decides how much autonomy is appropriate:

This is what “decision support” actually means in UniversitasAI. It’s not an AI looking at one number and triggering an action. It’s a seven-stage pipeline that involves deduplication, confidence scoring from historical data, cross-agent mesh consultation, policy enforcement, experimental integrity, and graduated human oversight — running thousands of times per day.

Layer 3: The Action Layer — Real Integrations, Not Mock Alerts

When the decision engine approves an action, UniversitasAI doesn’t just write a log entry. It executes through real-world integrations:

Layer 4: The Mesh Network — Agents That Talk to Each Other

The most architecturally complex layer. The 25 SphereAgents don’t operate in isolation — they communicate through an event-driven mesh network powered by the EventReactor, which processes 30+ event types and routes actions between agents automatically:

The Meta-Orchestrator sits above all 25 agents, running daily coordination cycles. It evaluates cross-department opportunities (e.g., “There are 40 students approaching graduation with incomplete career profiles — should Career Services and Student Success coordinate?”), dispatches low-risk actions automatically, and escalates complex multi-department decisions for human review. Importantly, managers and leaders can also be involved by proposing likely opportunities to be investigated — the system is not only autonomous but also a tool for leadership-driven exploration and what-if analysis.

The Complete Operational Loop

What Makes It Different

The 25 SphereAgents

Each SphereAgent is a domain expert AI that understands the rules, data, and priorities of its department.

The name “SphereAgent” reflects the architecture: each agent covers a sphere of institutional operations, and together they form a complete operational sphere around the entire university. They communicate with each other through a mesh network, share context, and coordinate actions — much like a well-run executive team, but operating 24/7 without fatigue.

How Decisions Are Made

AI with guardrails: a graduated system that gives AI more freedom for safe actions and more oversight for risky ones.

The central innovation of UniversitasAI is graduated autonomy. Unlike systems that are either fully automatic (dangerous) or fully manual (slow), UniversitasAI evaluates every proposed action and decides how much autonomy is appropriate. Crucially, the university itself can decide where any given type of decision should be placed across the four levels — the system provides intelligent defaults, but institutional leadership always has the final say on what level of AI autonomy they are comfortable with for each action type.

The Four Decision Levels

How Each Decision Is Made

Every proposed action flows through a confidence evaluation. Think of it like a credit score for AI decisions:

Approval Queue

When actions require human approval, they enter a priority queue with SLA tracking. Every escalated item has a deadline:

The Safety Architecture

Five layers of protection ensure that AI never takes an action it shouldn’t.

Trust is the foundation of any AI system in education. UniversitasAI implements five concentric safety layers that every proposed action must pass through. Think of them like airport security checkpoints — each one catches a different type of risk.

Agent Autonomy Modes

Administrators can dial each agent’s autonomy up or down independently:

The Emergency Stop

At any time, an administrator can press a single button to immediately halt all autonomous actions across the entire system. This is the outermost safety layer — it’s checked before anything else, and it overrides everything. Think of it as pulling the fire alarm: everything stops, instantly, system-wide.

Complete Audit Trail

Every single decision the system makes — whether auto-executed, escalated, approved, rejected, or denied — is recorded in an immutable audit log. This log includes who made the decision (AI or human), what action was taken, what confidence score was used, and what the outcome was. Nothing is ever deleted. This provides:

Learning & Self-Improvement

UniversitasAI doesn’t just execute actions — it measures their real-world impact and adjusts its behavior accordingly.

Most AI systems operate in an open loop: they take actions but never check if those actions worked. UniversitasAI has three independent learning mechanisms that create closed feedback loops:

Loop 1: Outcome-Based Calibration

Loop 2: Human Override Learning

Loop 3: Causal A/B Testing

Predictive Intelligence

The system doesn’t just react to problems — it sees them forming and intervenes before they become crises.

Compound Risk Scoring

A student with a 2.5 GPA might be fine. A student with a 2.5 GPA who also has overdue payments, declining class attendance, and no engagement with campus activities is almost certainly heading for dropout. UniversitasAI catches this compounding effect.

For every student, the system continuously computes a compound risk score from five independent signals:

Anomaly Detection

Beyond individual student risk, the system monitors institutional-level patterns using statistical analysis. If daily enrollment applications suddenly double, or payment volumes drop 40% below normal, or agent activity spikes unexpectedly — the system detects these anomalies automatically using Z-score analysis (measuring how many standard deviations a value is from its rolling average).

This catches situations that no pre-written rule anticipated — novel events, unexpected trends, and black swan situations.

Revenue Forecasting

The predictive engine decomposes revenue data into trend (are we growing?), seasonality (intake period spikes), and receivables (what’s outstanding and likely to be collected). This gives the CFO a forward-looking view with confidence bands, not just backward-looking reports. Leadership can also integrate future information and strategic plans — expected increases or decreases in new student enrollments, new revenue sources, planned program expansions — to create a comprehensive what-if analysis that blends AI prediction with human strategic insight.

Cross-Domain “Why” Analysis

When a KPI drops, administrators can ask “Why?” and the system queries up to four SphereAgents for their domain-specific perspective. The AI then synthesizes these perspectives into a ranked list of probable root causes with recommended interventions. What used to require a multi-department meeting now takes seconds.

The Stakeholder Experience

Students, faculty, and alumni don’t see the AI engine behind the scenes. They see a modern, engaging digital campus — each with their own tailored portal.

Student Portal

A full-featured progressive web app (installable on any phone, works offline) that gives students:

Faculty Portal

A dedicated portal for instructors and professors, purpose-built for academic workflows:

Alumni Portal

Graduates maintain a lifelong connection to ADSM through a self-service alumni portal:

Gamification & Digital Wallet

UniversitasAI makes academic engagement rewarding through a unique engagement-to-value pipeline:

Integration Ecosystem

UniversitasAI connects to the systems your institution already uses.

The platform is designed to work alongside existing tools, not replace them. It connects to enterprise resource planning (ERP), student information systems (SIS), learning management systems (LMS), payment processors, communication services, and document signing platforms through a unified integration layer. The system is also modular — you can take what you need, but you can also replace your existing systems entirely. It adapts to your institution’s technology landscape.

SIS Connectors

Pre-built connectors for the four major student information systems ensure UniversitasAI can work with any institution:

Integration Health Monitoring

All 12 external integrations are continuously monitored. If a provider goes down, the self-healing system automatically cycles through recovery strategies: retry, fallback, cache, circuit-break, and ultimately alert a human operator.

Financial Operations & ERP

A complete, built-in financial backbone — replacing the need for a separate ERP system like Odoo, SAP, or Oracle Financials.

Most universities run a patchwork of disconnected financial systems: one for accounting, another for procurement, a third for payroll, and yet another for student billing. UniversitasAI unifies all of these into a single platform where financial data flows naturally between operations. When a student pays tuition, the revenue is recorded in the General Ledger automatically. When a purchase order is approved, the budget allocation is updated in real time. When payroll runs, the salary journal entries post to the GL without manual intervention.

General Ledger & Accounting

The accounting module implements a full double-entry bookkeeping system with a UAE-specific Chart of Accounts pre-seeded with ~40 accounts organized by international accounting standards:

Every financial event produces a journal entry — a balanced pair of debits and credits that ensures the books always balance. The system enforces the fundamental accounting equation at the database level: no unbalanced entry can ever be saved.

Procurement & Inventory

The procurement module manages the complete purchase-to-pay lifecycle and tracks institutional assets and supplies:

Payroll & Compensation

A complete payroll engine built specifically for UAE labor law compliance. Handles the full monthly cycle from salary calculation to bank file generation:

The payroll lifecycle follows a controlled 4-step process:

Collaborative Intelligence

When agents think together: multi-agent deliberation, reasoning transparency, and coordinated action plans.

The first generation of UniversitasAI proved that specialized AI agents can automate institutional operations. This chapter describes the next evolution: agents that deliberate together, explain their reasoning, propose coordinated strategies, and learn from human feedback — transforming the system from independent workers into a collaborative intelligence.

Agent Council — Multi-Agent Deliberation

When a medium or high-risk situation is detected, the system automatically convenes the relevant agents to deliberate. Rather than a single agent’s recommendation, the decision-maker receives a multi-perspective assessment.

Agent selection is situation-aware. Student issues consult Student Success, Career Services, Registration, and Scheduling. Financial issues bring in Budget, Registration, and Student Success. The system maps each situation type to the perspectives that matter most.

Reasoning Chain Transparency

Every escalation now carries a structured reasoning chain — a step-by-step record showing how the recommendation was derived:

When a human rejects an action, they now select a structured reason (threshold too low, wrong entity, bad timing, insufficient context, policy violation). These structured rejections feed into the Human Override Learning engine, providing richer signals than free text alone.

Coordinated Action Plans

Instead of approving individual actions, agents can propose multi-step intervention plans that are approved and executed as a unit:

Ask ADSM — Unified Student Assistant

Students interact with a single conversational agent that orchestrates across all 25 SphereAgents behind the scenes. The student asks a question; the system classifies the intent, queries 2–3 relevant domain agents in parallel, and synthesizes a unified response. The student never needs to know which specialized agent answered — they just get a comprehensive, friendly answer with badges showing which domains were consulted.

Role-Based Briefings

The Executive Briefing extends to role-specific variants. Each role receives exactly the information relevant to their daily work:

Advanced Diagnostic Analytics

A 5-tab analytics dashboard fills the gap between descriptive (“what happened”) and predictive (“what will happen”) analytics, providing diagnostic insights: comparative KPIs with period-over-period trends, lead conversion funnel analysis, cohort retention matrices, behavioral student segmentation (4 clusters by GPA and standing), and an agent ROI heatmap showing financial impact by agent and tool. All zero-LLM-cost — pure SQL aggregation.

Security & Compliance

Built for trust: every layer designed with privacy, security, and regulatory compliance in mind.

Compliance Frameworks

Government & Accreditation Reporting

In addition to live connectivity with regulatory bodies like MOHESR, the system provides automated generation of compliance reports for three UAE regulatory bodies (CAA, MOHESR, ADEK) in their required formats — no manual data compilation needed. The system also prepares supplementary compliance reports, guidelines, and checklists for major international accreditation bodies including AACSB, EQUIS, and AMBA.

By The Numbers

The scale and depth of the platform today.

Platform Architecture

Production-Ready

The platform is deployed and operational with real institutional data, managing thousands of student records, leads, class schedules, and employee profiles in a production environment.

AI-Customized Learning Path Analytics

Personalized course recommendations based on learning style, academic performance, career goals, and labor market demand — moving from one-size-fits-all to genuinely adaptive education.

Traditional curriculum models assume every student learns the same way, progresses at the same pace, and enters the same job market. In reality, a Visual learner who absorbs content through video and diagrams has fundamentally different needs than a Reading-oriented student who thrives on documentation. A student targeting data science careers needs different electives than one aiming for corporate finance. UniversitasAI now analyzes these differences at scale and surfaces actionable insights for academic advisors and administrators.

Four Analytical Models

The Learning Path Analytics engine runs four complementary analyses across existing institutional data — no additional data collection required:

How Skill Gap Analysis Works

GPA-Adjusted Course Load

Not all students should take the same number of courses per semester. A student excelling academically can handle a heavier load, while a struggling student benefits from fewer courses with more support. The system calculates personalized recommendations:

Learning Style Classification

Rather than relying on self-reported preferences (which are often inaccurate), the system classifies learning styles from actual engagement data — what students do, not what they say:

The classification is deterministic and rule-based — no opaque ML clustering. Administrators can inspect exactly why each student received their profile classification. Students with insufficient engagement data are marked “unclassified” and excluded from the distribution.

Curriculum–Market Alignment

The most strategically valuable analysis answers: is our curriculum teaching what the job market actually demands?

The alignment score (matched / total demanded × 100) gives leadership a single number to track over time. A score of 68% means the curriculum covers about two-thirds of what employers are asking for — clear direction for curriculum committee discussions.

The Dashboard

All four analyses are presented in a 4-tab admin dashboard with interactive visualizations:

Student-Facing AI Learning Path

While the four admin-side models above serve institutional decision-making, students also need personalized, actionable guidance. The AI Learning Path feature gives each student a multi-term roadmap generated by gpt-5-mini, drawing from six data sources:

The LLM synthesizes all six inputs into a structured JSON response containing: recommended courses organized by term with priority labels (required/recommended/elective), skill milestones with current and target proficiency levels, career alignment scoring with gap areas, and a narrative summary. A rule-based fallback generates a reasonable path when the LLM is unavailable.

Key design decisions:

• Insert-only versioning — regeneration creates a new row and marks the old one is_current = False, avoiding expensive JSONB column updates and preserving path history for audit.
• 1-hour rate limit — Redis key with TTL (DB timestamp fallback when Redis unavailable) prevents excessive LLM calls.
• Auto-generate on first access — the GET /students/me/learning-path endpoint creates the first path automatically, removing friction.
• Autonomous scanner — paths older than 30 days are detected by the autonomous loop’s learning path scanner, which proposes low-risk regeneration to keep recommendations current with new course offerings.

Multi-Tenant Architecture

One platform instance serving unlimited institutions — each with its own data, branding, and configuration, fully isolated yet centrally managed.

The single most important architectural decision for commercial scalability: row-level multi-tenancy. Every table in the database (157 total) carries a tenant_id column that silently partitions data by institution. A student at University A can never see, query, or accidentally affect data belonging to University B — even though both institutions run on the same application server, the same database, and the same deployment.

How It Works

Infrastructure Isolation

Tenancy extends beyond the database:

Superadmin Management

A dedicated Tenant Administration dashboard (visible only to superadmins) provides complete institution lifecycle management:

Commercial Impact

Platform Observability

API metrics, event catalog, and custom report builder — the admin’s control panel.

As the platform grew to 180+ pages and 130+ API endpoints, administrators needed tools to understand system behavior, debug integrations, and generate custom data exports. Session 101 delivered three interconnected features that fill this gap.

API Usage Dashboard

Every API request is tracked through a lightweight middleware that writes fire-and-forget counters to Redis. A Celery task flushes these counters into the api_usage_hourly table every 5 minutes, giving admins a complete picture of request volume, latency, and error rates.

Path normalization replaces UUIDs and integers with {id} to prevent high cardinality — 1,000 student profile views collapse into a single /api/v1/students/{id} row rather than 1,000 separate entries. The dashboard offers 4 tabs:

• Overview: KPI cards (total requests, avg response time, error rate, unique endpoints) plus a hourly BarChart showing request volume and errors.
• Endpoints: Sortable table of top endpoints by request count, average latency, and error percentage. Click a column header to sort ascending/descending.
• Rate Limits: Configuration reference showing current rate limits for default API, auth endpoints, demo, bulk operations, and file uploads.
• Errors: Error breakdown with total count and a table of the top-20 failing endpoints sorted by error volume.

Webhook Event Catalog

UniversitasAI fires 24 distinct webhook event types across 8 categories (documents, registrations, leads, payments, students, enrollment, social, system). The Event Catalog is a 4th tab on the Webhook Management page that provides a browsable reference for each event type.

Each catalog entry includes the event type identifier, category badge, human-readable description, subscriber count (how many webhook endpoints listen for that event), last-fired timestamp, a sample JSON payload, and a field-by-field description table. Admins can filter by category, search by keyword, and expand any event to see its full payload structure.

Report Builder

The Report Builder provides a 5-step visual wizard for defining custom data exports:

Seven entity types are available (leads, students, registrations, documents, payments, employees, course sections), each with typed columns and filter definitions. Reports can be saved for reuse, and scheduled reports run automatically via a Celery cron task, delivering results by email to configured recipients.

Roadmap & Vision

Where we’re going next — and why it matters.

UniversitasAI is currently deployed and operational in a production environment. The platform has been built through 119 iterative development sessions, each adding new capabilities and hardening existing ones. Here is the forward-looking roadmap:

Session 119 delivered OBEF v11.5 compliance, parent portal, cascade workflows, adaptive thresholds, integration health monitoring, and dead letter queues — expanding the autonomous system to 106 scanners. The completed items below reflect this expansion:

Intellectual Property

The core innovations of UniversitasAI are protected by a provisional patent application (USPTO) covering nine foundational claims detailed in Chapter 1: graduated autonomous decision-making, multi-method institutional scanning, causal A/B experimentation, human override learning, compound multi-signal risk scoring, cross-agent mesh coordination, hybrid AI-mathematical optimization, adaptive institutional interface, and continuous environmental learning. The combination of these nine capabilities in a unified institutional platform is, to our knowledge, unprecedented in the EdTech and enterprise AI markets.

Additionally, the SphereAgent™ trademark registration is in progress, covering the brand name and the specialized agent architecture it represents.

Advanced Scheduling & Conflict Detection

Intelligent schedule validation with four conflict types, alternative suggestions, and visual analytics.

University scheduling is a constraint-satisfaction problem that grows exponentially with scale. Hundreds of course sections, limited rooms, instructor availability windows, minimum transition gaps between buildings — a single double-booking can cascade into dozens of disrupted classes. UniversitasAI now detects and prevents conflicts before they reach students.

Four-Type Conflict Detection

How It Works

When a high-severity conflict is detected during schedule creation or update, the API returns HTTP 409 with conflict details, preventing the change from being saved. Lower-severity conflicts are flagged but allowed. Administrators can also trigger a bulk scan across all schedules, persisting every detected conflict for review and resolution.

Alternative Suggestions

For each detected conflict, the system generates up to 8 conflict-free alternatives by scanning available time slots and rooms. Each suggestion includes the proposed day, start/end time, room, and a confidence score indicating how well it fits the existing schedule topology.

Visual Analytics

• Weekly Calendar View: Interactive time grid (7 AM – 9 PM, Mon – Fri) with color-coded course blocks. Hover for section details. This is the default tab on the Scheduling page.
• Peak Hours Heatmap: Class density visualization across all time slots, highlighting the busiest and quietest periods to guide schedule optimization.
• Instructor Workload: Per-instructor teaching hours, section counts, and schedule density for balanced assignment.

The Agent Nervous System

Real-time inter-agent communication that makes 25 SphereAgents behave like a living organism.

In the human body, the nervous system connects every organ through continuous signal transmission — a pain signal from the hand triggers a reflex in the arm, alerts the brain, and adjusts future behavior. UniversitasAI’s Agent Nervous System (ANS) applies the same principle to institutional AI: when one agent detects something significant, every relevant agent is notified instantly and can respond.

Signal Architecture

The ANS transmits five types of signals across four priority levels:

Cascade Chains

When a high-impact event occurs, the ANS triggers a cascade chain — a predefined sequence of agent activations that ensures comprehensive institutional response. Ten cascade chains are built in:

Shared Blackboard & Agent Pulse

Agents share state through a shared blackboard — a Redis-backed key-value store where any agent can write context that other agents read. For example, after detecting 15 at-risk students, Student Success writes at_risk_count: 15 to the blackboard; when Financial Aid runs its next cycle, it reads this count and prioritizes aid reviews accordingly.

Each agent also maintains a pulse — vital signs including health score (0–100), current status, signals emitted and received, and current focus area. The organism’s overall health is computed as the weighted average of all 25 agent pulses.

Visual Dashboard

The Nervous System page provides a full-screen interactive visualization built with React Flow:

• Network Graph: All 25 agents displayed as nodes in 7 domain clusters (Student Lifecycle, Enrollment, Academic, Operations, Engagement, Governance, Strategic Growth). Animated edges show live signal traffic between agents.
• Agent Detail Panel: Click any agent node to see its pulse vitals, pending signals, subscription list, and blackboard entries.
• Cascade Panel: Browse all 10 cascade chain definitions with step-by-step visualization. Trigger cascades manually for testing or emergency response.
• Signal Stream: Real-time feed of all signals flowing through the system, color-coded by priority (critical = red, high = orange, medium = blue, low = gray).
• Organism Vitals Bar: Global health score, active agent count, signals per minute, and active cascade count — always visible at the top.

Live Wire: Real Service Execution

As of Session 113, cascade steps are no longer abstract — they execute real backend service methods. All 52 steps across 10 chains are wired to actual services:

• Student Enrolled cascade: creates student success tracking record, awards welcome quest badge, initiates career profile, triggers schedule optimization
• Student At Risk cascade: triggers early warning alert, runs schedule conflict check, adjusts engagement scoring, notifies leadership
• Budget Crisis cascade: initiates hiring freeze review, pauses non-essential procurement, updates budget projections

Each action creates its own database session and is fail-soft — errors are logged but never block cascade propagation. Actions publish their results as SSE events so the dashboard shows success/failure in real time.

Server-Sent Events (SSE)

The ANS dashboard now receives live updates via Server-Sent Events. When a signal is emitted or a cascade propagates, the dashboard updates without polling:

• Signal events: New signals appear in the signal stream instantly
• Cascade events: Each cascade step is visualized as it executes (start → step → action result → complete)
• Keepalive pings: 30-second heartbeat prevents connection drops
• Auto-reconnect: 5-second delay on error, automatic cache invalidation on reconnect

The SSE endpoint authenticates via JWT query parameter (since EventSource doesn’t support Authorization headers) and uses Redis pub/sub with tenant-scoped channels for multi-institution isolation.

Executive KPI Dashboard

Chart-driven institutional analytics designed for deans, provosts, and university leadership.

University leaders need a single page that answers the question: “How is the institution performing right now?” The KPI Dashboard provides this through six integrated chart sections, each pulling from real institutional data with period filtering (6-month, 12-month, 24-month views).

Six Analytical Sections

Period Selector

A persistent 3-button toggle (6M / 12M / 24M) at the top of the page controls the time range for all charts simultaneously. Selecting a period re-fetches data with the appropriate date filter, enabling quick comparison across timeframes.

Dark Mode & Responsiveness

All charts use the platform’s shared useChartTheme() hook, ensuring correct axis colors, grid lines, and tooltip backgrounds in both light and dark mode. The layout adapts from a 2-column chart grid on desktop to single-column stacking on mobile.

Cross-Portal Intelligence

A unified intelligence layer that aggregates insights from students, faculty, and alumni into a single admin view.

Most universities have separate reports for each stakeholder group. Student success has its dashboards. Faculty workload lives in HR spreadsheets. Alumni outcomes are tracked in a separate CRM. The Intelligence Overview page collapses these silos into a single four-tab view that reveals cross-cutting patterns invisible to siloed teams.

Four-Tab Architecture

Actionable Insights

The overview tab surfaces cross-department insights that no single team would discover alone. Examples:

• “5 Faculty Overloaded” — HR + Scheduling: these instructors teach 5+ sections and have below-average office hour bookings. Suggests workload redistribution or additional section staffing.
• “18 Students Need Attention” — Student Success + Financial Aid: students with GPA below 2.0 who also have outstanding payment balances. Risk of dropout compounded by financial stress.

Zero LLM Cost

The entire Intelligence Overview is powered by pure SQL aggregation against existing database tables. No Azure OpenAI calls, no token costs, no latency. The CrossPortalIntelligenceService (280 lines) runs 5 queries across student, faculty, alumni, and office hours tables and computes all metrics server-side.

Student Intelligence Hub

AI-powered personal analytics giving every student a data-driven view of their academic journey.

Students traditionally have limited visibility into their own data. They see grades after the fact, discover financial issues when fees are overdue, and learn about graduation shortfalls in their final semester. The Student Intelligence Hub flips this model by proactively surfacing six categories of personalized analytics — all accessible from a single page in the student portal.

Six Intelligence Tabs

How It Works

Every metric is computed from the student’s actual database records via pure SQL — no LLM calls, no external APIs, no token costs. The StudentIntelligenceService (420 lines) runs per-query resilience: each tab’s data is fetched independently with its own try/except, so one failing query doesn’t break the entire page.

Privacy & Tone

The Intelligence Hub is designed with supportive, non-punitive language. Instead of “You are failing”, it says “Your GPA needs attention — here are steps to improve.” Risk signals use encouraging action verbs and link to relevant support resources. Peer benchmarks are anonymous and only appear when the cohort is large enough to prevent identification.

User Onboarding

A guided first-login experience that introduces every user to the platform’s capabilities — tailored by role.

A powerful platform is only useful if people know how to use it. UniversitasAI now includes a multi-step onboarding modal that appears on each user’s first login, introducing them to the features most relevant to their role. The onboarding is skippable, persistent (won’t reappear once completed or dismissed), and designed with the same semantic design tokens that power the rest of the interface.

Role-Specific Tours

Design Principles

• Non-blocking: Users can skip at any time. The modal won’t reappear once dismissed or completed.
• Persistent: Completion state is stored per portal in localStorage, so users who access multiple portals are onboarded to each one independently.
• Accessible: Full keyboard navigation (Escape to close, arrow keys to navigate, Enter to proceed). Screen reader compatible with proper ARIA attributes.
• Dark mode: Built with semantic design tokens — automatically adapts to the user’s theme preference.
• Demo-aware: Demo users (sandbox mode) skip onboarding entirely since they have their own guided experience.
• Mobile-responsive: The modal adapts to small screens with full viewport padding and touch-friendly button sizes.

Technical Implementation

The onboarding system is implemented as a single shared React component (OnboardingModal) in the shared-ui package, reused across all four portals with role-specific step configurations. The component renders via createPortal to avoid z-index stacking issues, uses CSS-only animations for step transitions, and checks localStorage on mount to determine first-login status — no backend changes required.

OBEF v11.5 Compliance

Automated scoring against MoHESR’s 24-KPI framework — the UAE’s new outcome-based evaluation standard for higher education.

The Outcome-Based Evaluation Framework (OBEF) v11.5 is MoHESR’s comprehensive quality assurance instrument for UAE higher education institutions. It defines how universities are measured, ranked, and funded. Manual compliance is a multi-month, error-prone process involving spreadsheets, survey coordination, and fragmented data sources. UniversitasAI automates the entire cycle — from data collection to scorecard generation to gap analysis to HEDB submission.

The Framework: 6 Pillars, 24 KPIs

What We Built

The OBEF compliance engine is a comprehensive subsystem: 9 database models, 27+ API endpoints, a normalization engine (Appendix A thresholds), an evidence vault (Appendix B audit trail), a survey sampling calculator (Appendix C), and an event compliance validator.

Future Readiness

A dedicated syllabus analyzer checks 8 skill categories against MoHESR’s future skills framework, positioning ADSM for the Future Readiness Label (≥90% skills coverage + ≥80% AI indicators). This is not a compliance checkbox — it’s a strategic advantage as MoHESR increasingly rewards institutions that prepare students for emerging technology roles.

5 OBEF Scanners

The autonomous loop includes five dedicated OBEF scanners that run every 10 minutes alongside the other 100 scanners:

The Parent Portal

Real-time visibility into your child’s academic life — grades, attendance, fees, and direct communication with faculty.

Parents are one of the most important stakeholders in education, yet most university systems treat them as outsiders. They call admissions for grade updates, visit campus to ask about fee balances, and learn about attendance issues weeks after they happen. The Parent Portal changes this by giving parents a secure, dedicated dashboard with real-time access to the information that matters most.

8 Pages, Purpose-Built

AI-Generated Weekly Report

Every week, the system compiles a personalized digest for each parent: attendance summary, grade changes, upcoming deadlines, and any flags from the student success system. This report is generated using pure data aggregation (zero LLM cost) and delivered both in-portal and optionally via email.

Design & Technology

• Dark mode: Full light/dark theme support using semantic design tokens, consistent with all other portals.
• Auth: Zustand store with parent-auth key, completely isolated from student/faculty/admin sessions.
• Shared UI: Built on the same shared-ui component library, ensuring visual consistency across the entire platform.
• Mobile-first: Responsive design optimized for the way parents actually check grades — on their phones.
• Multi-child: Parents with multiple enrolled children see all of them from a single dashboard with easy switching.

The Autonomous Hive Mind

106 scanners, self-training AI, 10 LLM providers, plagiarism detection, SAML SSO, natural language SQL, and more — how the system runs a school with minimum human intervention.

Chapters 2 and 4 introduced the autonomous loop and the 7-stage decision pipeline. This chapter documents what happened when we scaled that system from a prototype to a production-grade institutional nervous system: 106 scanners across every operational domain, 7 self-training feedback loops, 10 LLM providers with model benchmarking, plagiarism and AI content detection, SAML/ADFS SSO, natural language SQL queries, AI output monitoring, EDW connectors, recorded lecture interaction, practice exercise generation, and data retention automation.

106 Scanners: Full Domain Coverage

The scanner fleet has grown from the original 34 rule-based checks to a comprehensive 106-scanner system covering every operational domain:

Plus: 2 smart organization scanners, 7 intelligence-driven scanners, 10 support/social/ads scanners, 1 AI learning path scanner, 1 self-training scanner, 4 miscellaneous scanners, and 5 OBEF compliance scanners (Chapter 24).

Cascade Workflows

When a scanner detects a complex situation, a single action is rarely sufficient. Cascade workflows chain multiple steps with condition evaluation and shared context:

Each cascade step executes real backend service methods — these are not abstract workflows but wired integrations that trigger actual emails, create actual records, and update actual statuses.

Adaptive Policy Engine

The policy engine no longer uses static thresholds. It self-learns from 90-day outcome data:

Integration Health Monitor

16 integrations are checked every 30 minutes: Stripe, PayTabs, Adobe Sign, Azure Email, Telegram, WhatsApp, Odoo, Twitter, LinkedIn, Meta, and more. The monitor detects simulated mode (API keys empty), tracks 90-day availability history, and triggers critical alerts when a production integration goes offline.

Dead Letter Queue

Every failed Celery task is persisted to the database rather than silently discarded. The DLQ provides:

• Replay: Failed tasks can be retried with a single API call after the underlying issue is fixed.
• Health Summary: Dashboard showing failure rates by task type, critical task tracking, and trend analysis.
• Critical Alerts: Tasks marked as critical (payment processing, compliance submissions) trigger immediate notifications on failure.

Student Risk v2

The at-risk detection system now uses an 8-signal composite score: GPA trajectory, attendance trend, payment delays, course load, engagement level, social isolation, financial stress, and academic mismatch. When a student crosses the risk threshold, the system automatically suggests the appropriate cascade workflow.

Digital Twin v2

The institutional digital twin is now OBEF-aware. Administrators can run scenario simulations: “What if we improve retention by 5%?” and see the projected impact on OBEF scores. Recommendations are ROI-ranked, helping leadership prioritize investments that improve both outcomes and compliance scores.

Emergency Protocols

Crisis situations — mental health emergencies, campus safety incidents, Title IX reports — bypass the normal approval queue entirely. They enter a 1-hour SLA fast track with immediate admin email notification, counseling resource mobilization, and mandatory follow-up scheduling. These are the only actions that skip the 7-stage pipeline and execute immediately.

Self-Training AI: 7 Feedback Loops

The platform does not merely execute — it learns and improves itself through seven continuous feedback loops:

1. Prompt Evolution: Agent prompts are versioned and refined based on outcome quality metrics. Better prompts propagate automatically.
2. Scanner Threshold Adjustment: Detection thresholds self-calibrate based on 90-day false positive/negative rates.
3. ML Model Retraining: Enrollment prediction, retention forecasting, and risk scoring models retrain on new institutional data quarterly.
4. Embedding Refresh: Vector embeddings for agent memory and semantic search are periodically recomputed to reflect evolving institutional knowledge.
5. Cross-Agent Insight Sharing: Patterns discovered by one agent (e.g., admissions yield signals) are shared across the agent fleet.
6. Confidence Calibration: The OutcomeTracker adjusts agent confidence scores based on real-world action results.
7. A/B Experiment Promotion: Winning experiment variants are auto-promoted; losing variants are retired with statistical rigor.

Lead Intelligence

A Twin.so-style prospective student discovery engine monitors Reddit, GMAT forums, Quora, and education communities for high-intent candidates. The system detects intent signals (“looking for MBA programs in the Gulf”), scores lead quality, and routes discovered prospects to the CRM — before they ever fill a form.

10 LLM Providers & Model Benchmarking

The platform now supports 10 LLM providers: Azure OpenAI, OpenAI Direct, Anthropic, Google Gemini, AWS Bedrock, Mistral, Ollama Cloud, Together AI, Groq, and Fireworks. A model benchmarking system with 8 standardized tests evaluates each provider across accuracy, latency, cost, and education-specific tasks — enabling data-driven model selection per agent type.

Plagiarism & AI Content Detection

Integrated Copyleaks and GPTZero detection flags both traditional plagiarism and AI-generated content in student submissions. Academic integrity is built directly into the LMS workflow, with configurable thresholds, instructor review queues, and full audit trails.

SAML/ADFS SSO

Government identity federation via SAML and ADFS enables single sign-on for institutions with existing Active Directory or national identity providers. Critical for government and military institutions that require centralized authentication.

Natural Language SQL Agent

Staff can ask data questions in plain English: “How many students enrolled in MBA this semester?” The NL SQL agent translates queries to safe, read-only SQL, executes them, and returns visualized results with full query audit trail. No SQL knowledge required.

AI Output Monitoring & Content Policies

All AI-generated outputs are monitored against configurable content policies. The system flags inappropriate content, enforces institutional tone and terminology, and provides dashboards for compliance officers to review AI behavior patterns.

Additional Capabilities

• EDW Connector: Enterprise Data Warehouse integration for cross-system analytics and reporting.
• Recorded Lecture Interaction: Upload lecture recordings, auto-transcribe with AI, and enable students to ask questions against the transcript.
• Practice Exercise Generation: AI generates MCQs, flashcards, and exam prep materials from course content, personalized to learning objectives.
• Data Retention Automation: Policy-driven data lifecycle management with auto-archive, anonymization, and deletion schedules.

Leadership AI: The President’s Office Suite

Five new SphereAgents shipped in April 2026 target the decisions a university president actually makes — strategy, contested judgement calls, calendar discipline, hands-free command, and employee coaching — turning the Office of the President from a bottleneck into an amplified operator.

Chapters 3 and 4 introduced the 25 SphereAgents and the graduated autonomy system. The agents below do not replace operational decision-making; they give the president and senior leadership a permanent AI staff that works in the background and surfaces exactly the brief, recommendation, or prompt you need at the moment you need it.

Strategy Agent — Presidential Co-Pilot

The Strategy Agent is the home of institutional strategy. It holds the strategic plan, every KPI it maps to, every initiative that advances it, and the running reconciliation between planned and actual. Each morning it produces a one-screen brief: which objectives slipped overnight, which initiatives need a decision this week, which risks have newly entered the red band, and which wins can be celebrated today.

Delphi Decision Agent — Structured Consensus for Hard Calls

Not every decision is a one-person call. Faculty compensation bands, new-programme approvals, tuition changes, vendor selection, space allocation — these require structured input from multiple stakeholders. The Delphi Decision Agent runs the classic Delphi method as an asynchronous workflow: anonymous rounds, bounded convergence, audit-trail, and a final recommendation that leadership can adopt, modify, or over-rule.

Predictive Calendar — This Week, This Month, This Quarter

The Predictive Calendar does not show meetings. It shows priorities: the AI’s proposal of what this week, this month, and this quarter most need from the president. It scans every open cascade workflow, every escalated approval, every looming regulatory deadline, every slipping strategic initiative, every VIP touchpoint, and proposes where to spend the scarce resource: leadership attention.

• This week: 3–5 items requiring presidential input, ranked by consequence.
• This month: strategic decisions due, board materials to prepare, key external meetings to confirm.
• This quarter: milestones that, if missed, would threaten an annual goal. Early-warning view, not a to-do list.

Each item ships with the background context, the recommended option, and a one-click path to a full briefing.

Voice Interface — Hands-Free Command

For dean and admin use in meetings, cars, and between back-to-back events, the Voice Interface accepts natural-language voice commands and routes them through the same tool registry as the web UI. “Show me this week’s at-risk students”, “draft a note to the BoT about the Block B timeline”, “what’s Marc’s calendar tomorrow?” — all answered without unlocking a laptop. Destructive commands (“archive this record”, “cancel that meeting”) require an explicit verbal confirmation before execution, matching the graduated-autonomy model from Chapter 4.

Employee Twin — A Personalised HR Mirror

Every employee gets a twin: a dedicated agent instance that knows their role, their objectives, their completed trainings, their performance conversations, their known frustrations. Managers query the twin (“is Sarah on track for her Q2 objectives?”) and employees query their own twin (“what should I prioritise this week given my role goals?”). It is not surveillance; it is context held consistently on the employee’s behalf, with opt-in and a full audit of what the twin remembers.

Academic Intelligence for Faculty & Students

Six modules shipped in sessions 119–120 sit at the direct teaching–learning interface: lecture transcription & Q&A, AI-generated practice, structured peer review, quality-of-teaching evaluations, natural-language data queries, and AI output monitoring. Together they take what was already implicit (good teaching) and make it explicit (good teaching that compounds over time).

Lecture Transcript & Q&A

Faculty upload lecture recordings — audio, video, or a raw Zoom file. The system transcribes via Azure Speech / Whisper-class models, indexes the transcript alongside slides and readings, and exposes a chat surface where students ask questions against the lecture. “What did Dr Haddad say about the CAPM assumption at minute 22?” returns a timestamped quote and a jump link to the playback. Transcripts are searchable at the course and programme level.

Exercise Generator

The Exercise Generator produces quizzes, flashcards, scenario prompts, and short-answer practice sets from any course material in the knowledge base. Each item is tagged with the Bloom’s-taxonomy level it targets, the source chunk it came from, and the expected answer. Students self-test without waiting for faculty to author content; faculty review the generated items before publishing, retaining editorial control.

Course Evaluation Engine

Course evaluations stop being an end-of-term survey and become a continuous signal. The Course Evaluation Engine runs campaigns (surveys, reflection prompts, focus groups), aggregates responses, and surfaces actionable themes to instructors within 48 hours of fieldwork closing. Year-over-year comparisons are automatic; rubric-driven analysis flags specific courses where experience is drifting.

Faculty Peer Review

The Peer Review module supports the full assessment-of-assessment process that accreditors expect: paired reviewers, observation windows, rubric-driven comments, and calibration rounds. The agent nudges reviewers toward deadline, detects skew (reviewer A always grades gentler than reviewer B), and produces the package auditors ask for without the last-minute scramble.

Natural-Language SQL Agent

Chapter 26 mentioned this briefly; it deserves a fuller treatment because it changes who can answer data questions. The NL-SQL Agent translates plain English into parameterised SQL executed against a read-only view of the institutional database. Allow-listed tables, tenant-scoped rows, and automatic query-cost caps prevent the agent from returning data outside the caller’s authority. Every query (prompt, generated SQL, row count, latency) is logged for audit. Staff who would never write SQL can now ask “which programmes have the highest growth in applications this cycle?” and get a table and a chart in seconds.

AI Monitoring Dashboard

Every interaction between a user and an AI SphereAgent is logged as a structured record: prompt, response, tool invocations, token usage, latency, safety-classifier verdicts, and downstream database changes. The AI Monitoring dashboard surfaces patterns across these records — which agents are most used, which prompts fail classifier checks, which responses triggered corrective human overrides. It is the evidence base that lets a compliance officer answer “how is the institution using AI?” with data, not narrative.

Enterprise Safety, Operations & the Public Demo

An institution cannot trust an AI platform that is merely clever. It must also be demonstrably safe, demonstrably recoverable, demonstrably observable, and — when a prospective buyer clicks “try the demo” — demonstrably alive. This chapter documents the April 2026 hardening sweep that turned UniversitasAI into a defensible enterprise system.

Azure Content Safety — Model-Level Moderation

Every prompt sent to an LLM, and every response returned from one, is first passed through Azure Content Safety’s four-dimension classifier (Hate, Self-Harm, Sexual, Violence) at configurable severity thresholds. Blocked content is logged but never delivered to the user or persisted. When the Content Safety service is not configured — e.g. in local development — the system emits a loud structured warning and fails open explicitly, so the absence of a guard is always visible.

Azure Key Vault — Tiered Secret Resolution

Integration credentials, API keys, and webhook secrets are resolved through a three-tier chain: Azure Key Vault first (using a managed identity), then a tenant-specific encrypted DB row, then environment variables as a local-development fallback. The 5-minute in-memory cache keeps latency low. Secrets never sit in the container image and never appear in logs.

Backup & Disaster Recovery

The BackupService runs weekly (Sundays 02:30 UTC) and produces a gzip-compressed logical pg_dump of the full PostgreSQL database, uploaded to Azure Blob Storage with a SHA-256 checksum. Every backup is tracked on a BackupJob row with status, size, retention window, and a linkable RestorePoint. A daily retention sweep expires old jobs after the configured window (default 30 days). Restoration is a single CLI command against a target environment — tested quarterly, not left as an untested artefact.

Scheduled Production Smoke Tests

Every 30 minutes a Playwright suite runs against the live production URLs: backend health, every frontend origin, the demo-entry flow, a real monitor-account login, and three concurrent browser contexts exercising admin + student + faculty simultaneously. Failures upload trace and video artefacts and — when the Slack/Teams webhook secret is configured — alert the on-call channel within the 30-minute window, not hours later when someone notices.

16-Integration Health Dashboard

The Integration Health Service probes 16 external dependencies (Azure OpenAI, Odoo, Blackboard, Canvas, Moodle, Stripe, PayTabs, Tabby, Adobe Sign, Twilio, Meta, HEDB, MoHESR, ACS, LinkedIn, Telegram) and surfaces a status grid in the admin dashboard. Failed messages flow to a dead-letter queue with automatic retry and human-escalation paths. No silent failures.

Payment Status Awareness

Stripe and PayTabs clients expose a get_status() endpoint that reports live / test / simulated / not-configured. A dedicated admin banner warns when any payment processor is not in live mode, so demo fallbacks never silently reach production traffic.

Seven OBEF Scoring Models

Chapter 24 introduced OBEF v11.5 scoring. The engine now ships seven scoring strategies, not five: standard band, inverted band, binary, auto threshold, linear clamp, peer benchmark (scores relative to a cohort of comparable institutions), and longitudinal trend (scores rewarding improvement over time). The peer and trend models, added in April 2026, let the scorecard reward context and direction, not just level.

Adobe Sign Five-Template Registry

Document signing used to be an ad-hoc scattering of templates across services. The new registry centralises five named templates — offer letter, enrollment agreement, employment contract, NDA, scholarship agreement — each with typed signer roles, form fields, retention years, and emitted audit events. Integration code imports a template by name; swapping a template is a one-line change, not a cross-service refactor.

Ten LLM Providers

Cost and sovereignty vary by buyer. UniversitasAI now supports Azure OpenAI, OpenAI, Anthropic, Google Gemini, Ollama (local and cloud), Together, Groq, Fireworks, HuggingFace, and a custom OpenAI-compatible endpoint. The model benchmarking harness measures response quality, latency, and cost-per-call per provider, so the platform can recommend the right provider per task type — and institutions that require on-prem inference for regulated data can use Ollama-backed routing.

Lead Intelligence — Prospective-Student Discovery

Beyond managing inbound leads, the Lead Intelligence module actively discovers prospective students from public signals (LinkedIn graduates, conference attendees, scholarship announcements), scores them against the institution’s ideal-applicant profile, and suggests outreach campaigns. It is marketing reach normally only available to enterprise-scale admissions teams, delivered to boutique schools.

The Public Demo Tenant

Visitors to universitas.me can now click “Try Live Demo” and land inside a fully populated sandbox tenant within two seconds, no signup, no credit card. The demo tenant has its own deterministic UUID, is seeded on every boot with 100 students, 40 leads, 60 fee payments, 24 OBEF KPI scores, 12 employees, 4 departments, and 6 programmes. Demo writes are sandboxed by the DemoWriteProtectionMiddleware, session expires after two hours, and the tenant is completely isolated from real customer data — a prospective buyer can click through every page of the admin dashboard without ever seeing a real ADSM record.

The Demo-Defense Sprint — Making the Platform Argue for Itself

A platform that can do everything is not the same as a platform that shows everything. The April 2026 release made the system capable; the May 2026 sprint made the live demo at universitas.me argue for that capability in a CIO’s first sixty seconds. Three layers of work: surface the patent-bearing differentiator above the fold, animate the autonomy claim with live data, and lock the visible result behind blocking CI gates so it never silently regresses.

The First-60-Seconds Argument

A demo visitor lands on the dashboard and the first paint now communicates, in order: an AI Autonomy KPI tile showing the live count of autonomously executed decisions over the last seven days, the auto-execute rate, and the escalation count, with a USPTO 63/990,389 footer caption; a demo banner carrying both filed patent numbers (63/990,389 + 64/053,198) plus a “View as” persona switcher that mints a fresh JWT for the chosen role and opens the corresponding portal in a new tab; below the hero strip, a populated Approval Queue with six AI-proposed actions, each carrying a four-step reasoning chain, confidence score, and risk level. The previous demo showed an empty queue and a static autonomy widget. The new demo shows the patent-bearing claim made tactile.

The Live AI Activity Feed

The dashboard’s AI Activity ribbon was infrastructure-complete — an SSE stream endpoint, a card rendering recent decisions, a “LIVE” badge that lights up when connected — but had no source of new events during a typical evaluation window. The autonomous loop fires every ten minutes; a visitor watching the feed for ninety seconds saw nothing happen. The sprint added a per-minute synthetic AI pulse task that inserts plausible AIActionLog rows in the demo tenant only, rotating across ten sphere agents (student-success, scheduling, marketing, finance, hr, knowledge, career, ie, and others), mixing decision classes (auto_executed / confirmed / escalated), each tagged with a confidence score and decision reason. The SSE stream prepends new rows within three seconds; the LIVE badge pulses; the visitor sees the autonomy claim as observable behaviour rather than as marketing copy.

Layout Regression Defence in Three Layers

Mid-sprint a screenshot revealed truncation bugs in the dashboard chrome at common laptop viewports — tenant name clipping into the page area, KPI tiles overflowing the right edge at 1280-1599px, demo banner crowding the patent badge against the logout button. The fix shipped in three coordinated layers:

• Visible bugs fixed (Pass A). Sidebar header gets min-w-0 + truncate; hero KPI grid ramps 1 col → 2 cols → 4 cols at base / sm / xl breakpoints (was jumping straight from 2 to 4 at 768px); demo banner uses progressive collapse — full patent text on xl, abbreviated “Patent-pending” on lg, hidden below md.
• Shared primitives + design rules (Pass B). Three new shared-ui components — <TruncatedText>, <HeroKPIGrid>, <HeaderRow> — bake the four CSS rules that prevent the bug class (every grid child gets min-w-0; every text-bearing flex child gets min-w-0 + truncate; user-supplied content gets truncate; responsive grids declare every breakpoint). A new frontend/LAYOUT_RULES.md documents the underlying invariants for future contributors.
• Automated CI guard (Pass C). A new Playwright spec at e2e/tests/quality/layout-regression.spec.ts runs at three viewports (1280, 1440, 1920) across five admin pages, asserting (a) zero horizontal overflow and (b) every truncated element has a hover-discoverable title attribute. Plus nine pixel-level visual regression baselines committed under __screenshots__/ — any future PR that shifts the layout fails the build with a diff image.

Demo Tenant Must Produce Zero Real-World Side Effects

Three SLA-breach emails reached a personal address during the sprint — each from a seeded demo escalation whose deadline had naturally expired. The breach scanner was correct production code; it just had no concept of the demo tenant being non-operational. The fix codified a permanent rule: demo tenant is a display sandbox, not an operational tenant. Every external-effect call site now guards against demo-tenant work via a central app.core.demo_guard.skip_for_demo_tenant(tenant_id, action=...) helper. Nine surfaces were audited and protected:

Three independent failures would now have to align before any demo data could possibly reach a real-world destination. The rule is documented as gotcha #32 in CLAUDE.md so future contributors inherit it.

The Migration-Debt Story — Phantom 244 vs Actual 11

The new schema-drift CI gate caught what initially looked like a catastrophic finding: 244 tables declared on SQLAlchemy models had no formal Alembic migration. On a fresh database, alembic upgrade head produced zero tables despite reporting all 124 migrations as successful. Production was working only because Base.metadata.create_all() had been a silent backstop at app startup — load-bearing for over a hundred migrations.

The actual root cause was a single line in alembic/env.py. The async migration runner used connectable.connect() (non-autocommit) instead of connectable.begin(): every migration ran inside an implicit transaction that buffered until connection-close, then rolled back. Production hadn’t noticed because create_all() filled in everything that “migrations” were supposed to have done. After the one-line fix, real drift dropped from 244 to 11. Migration 125 (125_close_create_all_gap) closes that gap by deferring to Base.metadata.tables[name] for each missing table — idempotent on production, additive on fresh CI databases.

Quality Gates Now Blocking, Not Informational

Both quality systems that were “informational” coming out of April are now blocking, with severity tiers calibrated to fail loud on critical drift and log quietly on cosmetic variance:

• Schema-drift gate. Fails the build on missing tables in either direction or columns declared on models but not in migrations. Tolerates “extra columns” in migrations as INFO-only (production has them, harmless until someone writes an explicit cleanup migration).
• Tenant-mixin invariants. Three blocking unit tests asserting every non-global model uses TenantMixin, every tenant_id column is NOT NULL with no default, and the audit allowlist is empty — StrategyRun and PredictionRun retired into GLOBAL_TABLES with their docstring rationale.
• Layout regression. Playwright spec at three viewports + pixel-level baselines for five admin pages.
• External-effect demo guards. Verified by inspection across nine surfaces; future audits run through the same central helper.

Sales-Tour Screencast for Board-Pack Replay

Captured at docs/marketing/sales-tour-2026-05-01.webm — a 90-second walkthrough of the dashboard hero, AI Activity feed, Approval Queue, OBEF Scorecard, role switcher, and Ask-the-dashboard panel. 2.2MB, plays natively in every modern browser, regenerable on demand via node tools/sales-tour-screencast.mjs. The demo now has both a live experience for evaluators and a replayable artefact for board packs.