01
Intake
Scanned PDFs, lab faxes, and EHR exports land in S3 via signed-URL uploads with a typed audit log.
MVP · 7-day plan
auto-med-indexer
An automated indexer for medical records.
Seven working days from a blank repository to a deployed MVP: ingest scanned records, extract clinical fields, and surface them through a searchable dashboard that survives a clinician's first hour on the job.
Week at a glance
The plan
One deliverable per day, in order.
Each day stands on its own. Cancel any day and the code already shipped still builds and still runs. The plan is short on purpose - seven days, four bullets each, no goals beyond the working artifact.
- 01
Repository scaffold and core routing
Stand up the TanStack Start app at the repo root, define the route surface, and land a clean CI build.
- Initialize the TanStack Start app at the repo root and lock the build target
- Define the route surface - "/" plus the /api/* family used by the rest of the plan
- Set up shared types, env scaffolding, and a conventions note for contributors
- Land a green CI build and a previewable shell before any feature work
- 02
S3 document ingestion
Wire the storage layer so uploaded scans and EHR exports land in S3 with a typed audit trail.
- Configure the S3 bucket, prefixes, and signed-URL handlers for direct browser uploads
- Implement POST /api/uploads for server-side signing and presigned PUT URLs
- Build the upload UI with progress, retry, and MIME validation for scans and PDFs
- Persist the storage audit log - upload id, original filename, owner, content type
- 03
OCR and medical-field extraction, server-side
Add the server-side pipeline that turns raw scans into structured records with clinical codes, focused on radiology and imaging studies.
- Server function: pull from S3, run OCR, normalize layout and whitespace
- Extract radiology fields - patient id, encounter date, document type, modality, body part, ordering provider
- Tag records with ICD-10 and SNOMED CT codes, each with a confidence score
- Write the index record alongside the source document with a deterministic join key
- 04
Search and filtering
Surface the indexed corpus through a query layer that clinicians can actually drive.
- Search endpoint with full-text plus structured filters (date range, code, patient id)
- Paginated results, sortable by date, document type, or code frequency
- Highlight the matching token in the result snippet so reviewers see why it landed
- Honest empty-state copy that names the active filters and explains no hits why
- 05
User management and clinical dashboard
Put a role-aware shell around the index so each user only sees what their role allows.
- Users and roles table - clinician, admin, auditor - with server-side authorization
- Dashboard cards: document totals by system, recent uploads, top tagged codes
- Per-document detail view with extracted fields, code tags, and the source citation
- Server-side authorization check on every read; never trust the route
- 06
Final testing and edge case handling
Cover the cases the happy path would have missed and lock the contract behind tests.
- Cover broken-OCR, oversized scans, expired auth, and malformed uploads with typed error codes
- Add retry-with-backoff to the indexing server function and cap its maximum runtime
- Snapshot and unit tests for the search response shape, the pipeline round-trip, and the lexicon
- Manual walkthrough of the fixture corpus before sign-off
- 07
Deployment and handover
Cut over to production and leave behind a clean repo, a one-page runbook, and a verified health probe.
- Production build, environment wiring, and a verified health probe on /api/health
- One-page handover doc - runbooks, owner, escalation path, on-call rotation
- Final repo cleanup: no scratch files, no debug printouts, no leftover fixtures
- Verify every route at the live URL before handing the keys back
Architecture
The four-stage pipeline.
Every stage is a server function with a single responsibility. The stages run sequentially when a new document arrives and incrementally as the corpus grows, with a deterministic join key tying each index record back to its source object.
02
Extract
A server function pulls the object, runs OCR, and structures the text into a flat record of clinical fields.
03
Index
Records are tagged with ICD-10, SNOMED CT, and a patient identifier, then written to the index table.
04
Retrieve
The dashboard queries by date, code, or patient id and returns a snippet linked back to the source document.
Status
7 / 7 · MVP shipped