ClaimFlow AI

ClaimFlow AI is a document-intake workflow for insurance claims.

It lets a user upload a claim PDF or paste claim email text, extracts structured claim JSON using Gemini, validates the extracted data with deterministic rules, saves the run in Postgres, and shows a traceable timeline.

Current Scope

Supported inputs:

• PDF claim document
• Pasted email text

Current output:

• Structured claim JSON
• Model and prompt version
• Confidence JSON
• Missing fields
• Conflicts and warnings
• Required evidence
• Final status: COMPLETED or NEEDS_REVIEW
• Saved extraction run
• Timeline events

Tech Stack

• Turborepo
• Next.js
• TypeScript
• Prisma
• Postgres
• Zod
• Gemini API
• Zustand

Workflow

PDF / email text
→ Document row
→ ExtractionRun row
→ DOCUMENT_UPLOADED event
→ Run extraction
→ Gemini extraction
→ Zod-parsed structured JSON
→ extractedJson saved in Postgres
→ status moves to VALIDATING
→ Run validation
→ deterministic rules check missing fields, conflicts, warnings, and evidence
→ validationJson + missingFieldsJson saved in Postgres
→ status becomes COMPLETED or NEEDS_REVIEW
→ validation result + timeline shown in UI

Progress

Day 1: Foundation

Implemented:

• Turborepo setup
• Next.js app
• Prisma + Postgres
• Shared Zod schemas
• DB models:
  • documents
  • extraction_runs
  • extraction_events

Day 2: Upload + Run Creation

Implemented:

• PDF upload
• Email text submission
• Local PDF storage
• POST /api/documents/upload
• Run creation with UPLOADED status
• Dashboard recent runs list
• Run detail timeline

Day 3: Gemini Extraction

Implemented:

• @repo/ai package
• Gemini client
• Prompt versioning
• PDF extraction
• Email text extraction
• POST /api/extraction-runs/[runId]/extract
• Manual Run extraction button
• Extracted JSON panel
• Timeline events:
  • EXTRACTION_STARTED
  • MODEL_RESPONSE_RECEIVED
  • EXTRACTION_COMPLETED
  • RUN_FAILED

Day 4: Validation Layer

Implemented:

• Deterministic claim validation after extraction
• validateClaimExtraction() shared validation service
• POST /api/extraction-runs/[runId]/validate
• Manual Run validation button
• Missing fields detection
• Conflicts and warnings detection
• Required evidence detection
• Final status:
  • COMPLETED
  • NEEDS_REVIEW
• Saved validationJson and missingFieldsJson
• Validation summary UI
• Missing fields UI
• Conflicts, warnings, and required evidence UI
• Timeline events:
  • VALIDATION_STARTED
  • VALIDATION_COMPLETED
  • MISSING_FIELDS_DETECTED
  • CONFLICTS_DETECTED
  • RUN_COMPLETED
  • RUN_NEEDS_REVIEW

Day 4 flow:

Upload PDF or paste email text
→ created run
→ click Run extraction
→ Gemini extracts structured claim JSON
→ app saves extractedJson
→ status moves to VALIDATING
→ click Run validation
→ app checks missing fields, conflicts, warnings, and required evidence
→ app saves validationJson and missingFieldsJson
→ status becomes COMPLETED or NEEDS_REVIEW
→ UI explains the result
→ timeline proves what happened

Tested with synthetic PDFs:

• valid own-damage claim → COMPLETED
• missing policy number → NEEDS_REVIEW
• repair estimate only → NEEDS_REVIEW
• third-party without police report → NEEDS_REVIEW
• theft claim missing FIR → NEEDS_REVIEW

Current Run Lifecycle

Completed lifecycle:

UPLOADED
→ EXTRACTING
→ VALIDATING
→ COMPLETED

Needs-review lifecycle:

UPLOADED
→ EXTRACTING
→ VALIDATING
→ NEEDS_REVIEW

Failure lifecycle:

UPLOADED / FAILED
→ EXTRACTING / VALIDATING
→ FAILED

Environment Variables

DATABASE_URL=postgresql://postgres:postgres@localhost:5432/claimflow_ai?sslmode=disable
GEMINI_API_KEY=
GEMINI_MODEL=gemini-2.5-flash

For local extraction, keep Gemini variables in:

apps/web/.env.local

Run Locally

bun install
bun run db:migrate
bun run dev

App runs at:

http://localhost:3001

Next Step

Week 2:

• Add human review queue
• Show runs that need review
• Allow reviewer decisions
• Store reviewer notes and corrections
• Use reviewer feedback later for memory/rule improvement