Fornix AI
An AI-powered platform that turns scattered, paper-bound school incident reports into a single source of truth — and into foresight, FERPA-compliant by design.
What this project was for
Arrived 4–5 months into a stalled build — only file upload existed, no architecture. We turned it around: AI vision auto-extracts paper incident reports, ML forecasts risk before it escalates, FERPA-compliant by design.
The problem
School districts are responsible for thousands of students across dozens of campuses, yet the data that should inform safety decisions is fragmented. Incident reports arrive as PDFs, scanned forms, and paper, then get re-keyed by hand into disconnected spreadsheets. Districts end up data-rich but insight-poor: crushing manual entry, no district-wide visibility, reactive rather than proactive, and real FERPA exposure.
How we delivered
We treated it as a turnaround — stabilize first, then build fast on a foundation we could trust. We audited the codebase, defined the architecture, schema, and AI pipeline up front, then time-boxed the remaining scope (document processing → analytics → forecasting → reporting → administration). The AI/ML work was a dedicated testable service layer, with FERPA compliance and audit logging engineered in from the start, shipped to a monitored CI/CD cloud environment.
What we built
AI document processing
an AWS Bedrock vision model reads each PDF/scan and extracts structured incident data; multi-page PDFs split into linked records; PII stripped during processing; source files auto-deleted; smart school-name matching and duplicate detection.
Insight and foresight
a centralized cross-district dashboard, ML forecasting of incident volumes/types/costs (calendar-aware), geographic hotspot analytics, and one-click board-ready reporting.
Enterprise access & trust
role-based access across districts/schools/staff, Microsoft SSO, and a full audit trail logging every access with user, IP, and device.
Built to run in production
Clean data from messy documents
vision extraction + alias maps, fuzzy matching, and duplicate detection.
Privacy & FERPA
PII anonymized during processing, source files auto-deleted, complete attributable audit trail.
Heavy workloads without blocking
extraction offloaded to Celery + Redis with progress tracking, retries, exponential backoff.
Accurate, usable analytics
Prophet / XGBoost / scikit-learn combined with domain-aware logic (weekends, holidays) and clear interactive charts.
Results
A stalled, partial build delivered as a complete, production-ready platform on a committed timeline.
Manual data entry eliminated; district-wide visibility consolidated into one dashboard.
Proactive risk management via ML forecasting and geographic hotspot analytics.
Privacy and compliance engineered in by design, with a full audit trail.
How it was built
We took the project over as a rescue: audit first, then a defined architecture, schema, and AI pipeline before writing feature code. The remaining scope was time-boxed in a clear sequence — document processing, analytics, forecasting, reporting, then administration.
The AI/ML work lives in a dedicated, testable service layer rather than being scattered through the app. An AWS Bedrock vision model reads each PDF or scan and extracts structured incident data; heavy extraction is offloaded to Celery + Redis with progress tracking, retries, and exponential backoff so the interface never blocks.
Compliance was engineered in from the start: PII is stripped during processing, source files are auto-deleted, and every access is logged with user, IP, and device for a complete audit trail. Forecasting combines Prophet, XGBoost, and scikit-learn with calendar-aware domain logic for weekends and holidays.
Vision document pipeline
AWS Bedrock reads PDFs/scans, splits multi-page reports into linked records, and matches school names.
Non-blocking heavy work
Extraction offloaded to Celery + Redis with progress tracking, retries, and exponential backoff.
FERPA by design
PII anonymized during processing, source files auto-deleted, and a full attributable audit trail.
Calendar-aware forecasting
Prophet / XGBoost / scikit-learn combined with domain logic for weekends and holidays.
What it runs on
Need a production system like this?
Tell us what you're building, fixing, automating, or scaling. Book a technical call to review scope, architecture, and timeline — or get an estimate and we'll come back with honest feedback on feasibility and approach.