Document Intelligence for Insurance Claims

AI-assisted back-office tools for insurance claims workflows. Concept demo using UNIQA-style branding. Not affiliated with or endorsed by UNIQA Insurance Group — used illustratively for portfolio purposes.

🟢 Live demo: https://document-intelligence-poc.onrender.com

First load may take ~30 seconds — Render free tier sleeps the container after 15 minutes of inactivity.

---

The business problem

A typical claims adjuster in an insurance back-office spends 5–15 minutes per case re-typing data from incoming PDFs and emails into the internal claims system (TIA, Guidewire, custom systems). At 25 cases per day across a 30-person team, that's ~75 hours of manual data entry daily — done by qualified domain experts who could instead be reviewing complex claims, talking to customers, or training junior staff.

This demo shows three AI-assisted tools that move that bottleneck:

Tool	Input	Output	Time saved per case
Extract	PDF claim form / invoice	Structured JSON, fields verified against PDF	~8 minutes
Classify	Incoming email / message	Category + priority + routing team + actions	~3 minutes
Summarize	Long narrative / email thread	Executive summary + key facts + action + reply draft	~10 minutes

The adjuster stays in the loop on every case: the AI proposes, the adjuster confirms (or corrects) before the result is pushed downstream. Missing or low-confidence fields are surfaced for review in amber.

---

60-second self-guided tour

Open the live demo and try:

1. Click "Sample claim form" — a Slovak car-insurance claim form loads. Watch the PDF preview render on the left while structured fields auto-populate on the right.
2. Click any field on the right (e.g. Policy number, Total, Bank) — on the PDF preview, the source word(s) get highlighted in yellow. Every extracted value can be traced back to its origin on the original document.
3. Switch to Classify mode (header nav) and click Payment complaint — a customer threatening to escalate to the Slovak National Bank loads.
4. Click Classify case — notice the 🔴 High priority badge (the model recognises the regulatory escalation language) and the routing to the Billing & Reimbursement team with concrete suggested actions.
5. Click Approve & route — a success card shows the case ID, the team mailbox, the audit log, and (if a webhook is configured) confirmation that the payload was actually delivered to a downstream system.
6. Switch to Summarize mode and click Disputed claim thread — a multi-message email argument loads. Click Summarize to get a 2-3 sentence executive summary, the key facts, a detected sentiment (😠 angry), a suggested action, and a ready-to-edit customer reply.
7. Watch the Time saved counter in the header grow with each action.

---

Architecture

flowchart LR
    A["📄 PDF (Extract)"] --> B["pdfplumber<br/>text + word bboxes"]
    A2["📧 text (Classify)"] --> C
    A3["📝 long text (Summarize)"] --> C
    B --> C{"ANTHROPIC_API_KEY<br/>configured?"}
    C -- "yes" --> D["Claude Opus 4.7<br/>structured outputs"]
    C -- "no" --> E["Regex / keyword<br/>heuristic fallback"]
    D --> F["Pydantic validation"]
    E --> F
    F --> G["Review UI<br/>(form / cards / preview)"]
    G -- "approve" --> H["POST /integrate"]
    H --> I{"INTEGRATION_<br/>WEBHOOK_URL<br/>set?"}
    I -- "yes" --> J["📡 Live POST<br/>to webhook"]
    I -- "no" --> K["Mock success card<br/>(case ID + email + audit)"]

Loading

Two run modes — both visible in the UI

Mode	When	Visual indicator
LLM (claude-opus-4-7)	ANTHROPIC_API_KEY is set	Green "claude-opus-4-7 extraction" badge
Heuristic (regex)	No API key	Amber "Heuristic extraction (no API key)" badge

The heuristic fallback means the demo works end-to-end without any API credentials — anyone can clone the repo and see the full flow. Recruiters clicking the live URL see the heuristic mode by default; the LLM mode unlocks when an API key is added.

Two integration modes — both visible in the UI

Mode	When	Visual indicator
Live webhook	INTEGRATION_WEBHOOK_URL is set	Green "LIVE WEBHOOK" badge + HTTP status
Mock	No webhook URL	Amber "MOCK" badge — case ID + email are fake

---

What's interesting technically

1. Click-to-highlight source verification — pdfplumber exposes per-word bounding boxes; the frontend renders the PDF via pdf.js and overlays a highlight canvas. Clicking a field in the form finds the matching words on the page and draws yellow boxes around them. This is the "decision-support" loop in a single interaction.
2. Strict-schema structured outputs from Claude — uses output_config.format: json_schema so the model is constrained to the exact Pydantic model the rest of the pipeline expects. No prompt-engineering hacks to "please return JSON" — the API enforces it.
3. Graceful degradation — every AI call has a deterministic regex fallback. The product works without external dependencies and the UI is honest about which mode is active.
4. Three tools, one workspace — header nav switches between Extract, Classify, and Summarize via URL hash (#extract / #classify / #summarize). Same backend, same styling, shared time-saved counter, shared /integrate downstream path.
5. Real vs mock integration toggle — toggling INTEGRATION_WEBHOOK_URL moves the same Approve flow from mock to live POST without code change. Lets you demo the same UI as either "works without setup" or "production-style pipeline".

---

Run locally

# 1. Install deps (once)
python -m pip install -r requirements.txt

# 2. (Optional) Add API key for the LLM path
copy .env.example .env
# edit .env and set ANTHROPIC_API_KEY=sk-ant-...

# 3. Start the server
.\run.ps1
# or:
python -m uvicorn app.main:app --reload --app-dir .

Open http://127.0.0.1:8000.

---

Deploy

The repo ships with a Dockerfile and a render.yaml blueprint. To deploy your own copy on Render's free tier:

1. Fork this repo on GitHub.
2. On https://render.com, New → Blueprint and point at the fork.
3. Render reads render.yaml. Add ANTHROPIC_API_KEY as a secret env var if you want the LLM path — leave empty to deploy in heuristic mode.
4. (Optional) Add INTEGRATION_WEBHOOK_URL for the live downstream integration demo — see next section.
5. First deploy takes ~3 minutes.

The Docker image runs unchanged on Fly.io, Railway, Google Cloud Run, or any container host.

Adding the real downstream integration

For interviews it's powerful to show data actually flowing somewhere — not just appearing on screen. Set this up once:

1. Open https://webhook.site in a new tab. It auto-generates a unique URL.
2. In Render → service → Environment, add INTEGRATION_WEBHOOK_URL with that webhook.site URL as the value. Save.
3. Render redeploys automatically (~30s).
4. Open the demo and the webhook.site tab side by side. Click Approve & route or Save & export — the payload appears in webhook.site in real time, the result card shows a green "LIVE WEBHOOK" badge with the HTTP status code.

When the env var is not set, both buttons still work — they just show a "MOCK" badge instead.

---

Project layout

app/
  main.py        FastAPI app, routes, static serving
  extractor.py   Extract orchestration (PDF → LLM or heuristic)
  pdf_reader.py  pdfplumber wrapper, text + per-word bounding boxes
  llm.py         Claude client, JSON schema, heuristic fallback
  classifier.py  Classify orchestration (text → LLM or heuristic)
  summarizer.py  Summarize orchestration (text → LLM or extractive heuristic)
  integration.py Mock + real downstream integration (webhook POST)
  samples.py     Built-in sample messages + claim narratives
  schemas.py     Pydantic models
static/
  index.html     UI shell with mode nav (Extract / Classify / Summarize)
  app.js         Frontend logic + PDF.js + click-to-highlight
  favicon.svg    Document-mark favicon
  style.css      Custom styling
sample-data/
  invoice.pdf      Generic invoice
  claim-form.pdf   Slovak car-insurance claim form (generated)
scripts/
  generate_claim_sample.py  fpdf2 generator for the claim form
  generate_banner.py        README banner generator
  generate_og_image.py      Social-share card generator
  smoke_test.py             End-to-end smoke test against a running server
tests/
  test_extraction.py     Heuristic extraction + sample-PDF assertions
  test_classification.py Category / priority / routing assertions
  test_summarization.py  Summary / sentiment / draft assertions
  test_integration.py    Downstream routing + case-ID format
.github/workflows/ci.yml  Runs the test suite on every push
Dockerfile  · render.yaml  · run.ps1

Tests

The pure-Python logic (heuristic extraction, classification, summarization, routing) is covered by a pytest suite that needs no API key:

pip install -r requirements-dev.txt
python -m pytest tests/ -q

GitHub Actions runs it on every push — see the CI badge at the top.

---

Roadmap to production

What this demo deliberately does not include — and what would need to be added for a real deployment:

• Authentication (SSO / OIDC against the company directory) and per-adjuster audit log
• Persisted extractions + reviewer corrections — corrections are high-quality training data for fine-tuning
• LLM-returned per-field confidence scores (not just a binary missing flag) — needed for "auto-approve if all fields >95%" workflows
• Batch processing via the Claude Batches API for end-of-day bulk runs
• Real downstream integration with a specific claims system (TIA, Guidewire, custom) — currently demonstrated via a configurable webhook URL
• Replace heuristic fallback with a model-distilled small classifier so the no-API-key mode has higher fidelity
• Multi-page PDF support is implemented; multi-document batch upload UI would be ~1 day

---

License

MIT — see LICENSE. The UNIQA name, logo, and brand colours are property of UNIQA Insurance Group. This is an independent portfolio project not affiliated with UNIQA.