Control
Human-owned decisions
Approval, rejection, fraud, pricing and eligibility decisions are explicitly blocked from model execution.
AI engineering portfolio · synthetic data only
checking runtime…A bounded insurance knowledge assistant demonstrating document extraction, retrieval-grounded generation, PII controls, human decision boundaries, persistent serverless retrieval and AWS AI integration.
Step 1
The public path is intentionally synthetic-only. When the AWS backend is connected with Textract enabled, the bundled PDF is processed through Textract; the fallback path uses pre-bundled synthetic text and states that clearly.
No document loaded yet.
Only use fabricated or anonymised files. Public deployments disable this route by default.
Step 2
Answer
Control
Approval, rejection, fraud, pricing and eligibility decisions are explicitly blocked from model execution.
Privacy
PII is redacted before indexing/model use and output is checked again before returning to the client.
Reliability
DynamoDB persists document chunks across Lambda cold starts; OpenSearch vector retrieval remains an optional scale-up path.
Governance
Structured request metadata goes to logs while raw documents, prompts and evidence are deliberately omitted.
Deployment
The interface exposes whether requests are hitting AWS Bedrock or the synthetic Cloudflare fallback, preventing portfolio overclaiming.
Verification
Backend tests, evaluation cases, JS syntax checks and edge smoke tests cover retrieval, policy boundaries and deployment behaviour.
Architecture
Browser ↓ Cloudflare Pages + same-origin Function ↓ API Gateway ↓ Lambda + FastAPI ├─ Bedrock Converse ├─ DynamoDB document chunks ├─ Textract optional OCR └─ CloudWatch structured audit metadata
The edge fallback exists only to keep the public synthetic demo usable before AWS credentials are connected. It is deliberately labelled in the live runtime status above.