A records-review platform with an AI extraction pipeline
An AI-assisted, physician-reviewed medico-legal record review platform — from secure record intake through automated analysis to a source-cited, verified deliverable.
The problem
Personal-injury and malpractice firms work cases with thousands of pages of disorganized medical records — duplicate scans, missing date ranges, handwritten notes, mixed providers. Building a chronology by hand is slow, expensive and hard to defend once a conclusion loses its source page — but a raw LLM summary isn't defensible either.
- A case arrives as thousands of pages — duplicate scans, missing date ranges, handwritten notes, mixed providers
- Building the chronology and finding the decisive fact is slow and expensive by hand
- A conclusion that loses its source page is hard to defend in a deposition
- A raw LLM summary is faster, but it has no citations and can invent facts
The approach
- An async pipeline runs OCR and extraction into a strict schema — the model fills fields, it can't invent them
- Every extracted value and every search result carries its source page
- RBAC is enforced per-endpoint on the API; the front-end route guard is only a UX convenience
- A physician reviews and signs off before any deliverable is released
What we built
Secure intake wizard
Multi-step case creation with client-side validation, draft-and-resume, and encrypted upload to object storage.
Async AI document pipeline
Digital PDFs are parsed directly; scanned pages fall back to a vision model for OCR and extraction in one pass, validated against a strict schema.
Source-cited semantic search
Natural-language questions across a firm's whole corpus, returning ranked snippets that link back to the exact case and file.
Per-case AI summary
A generated case narrative stored with the case, with the model name recorded for auditability.
Reviewer workflow
Physicians and reviewers see only their assigned cases, update status, and upload verified deliverables.
Billing & audit trail
Consolidated multi-case invoices, sequential numbering under a DB transaction, payment-gated downloads, and an immutable per-case event log.
How it works
- 01
Intake
The firm creates a matter, selects services, and uploads records through the guided wizard.
- 02
Extract
Each file is queued to a worker: text or vision extraction into a validated schema, then chunk-and-embed for search.
- 03
Review
A physician checks the extraction and the generated summary, corrects anything, and signs off.
- 04
Deliver
The verified deliverable is released once the invoice is paid; every step lands on the case timeline.
Architecture
Engineering
Server-side RBAC by default
A global auth + roles guard protects every route; endpoints opt out explicitly. Four roles — client (firm-scoped, team-seat visibility), admin, reviewer, billing.
Defense-in-depth vector search
Qdrant payload filters scope results for speed, but every hit is re-checked against the caller's real RBAC scope in application code before it's returned.
Resilient job pipeline
Embedding is best-effort — a search-index failure doesn't fail the user-visible extraction. Re-running extraction clears stale vectors first.
CI quality gate
Policy checks, lint, typecheck, build, dependency audit, and a Docker build + Trivy scan (HIGH/CRITICAL fails the build) for both apps.
By the numbers
Target turnaround per matter
Accuracy goal, with mandatory human review
Encryption in transit and at rest, least-privilege, audit logging
The product's stated targets — mandatory human review before every delivery.
Outcome
- Manual re-keying from records dropped to near zero
- Every search result and extraction carries its source page — defensible, not a black box
- Clients track case status themselves; every deliverable ties back to an approval and a payment
Stack
Let's build what comes next.
Have an idea, product or technology challenge? Let's talk about what we can build together.