Flagship applied AI system
Governed agentic AI workflow for motor-insurance claims.
ClaimFlow AI turns unstructured claim PDFs and emails into structured, policy-grounded, human-reviewed cases. It combines extraction, deterministic validation, clause-level policy RAG, guarded agent tools, workflow memory, evals, and run-level observability so the AI can assist without owning the final claim decision.
Claim processing is risky to automate directly because claims are incomplete, policy-dependent, and require accountability. A basic LLM can extract text or generate answers, but it cannot safely own missing-field validation, policy evidence, workflow routing, or final decisions.
ClaimFlow AI turns the claim into a stateful workflow. It extracts structured data, validates missing information, retrieves policy clauses, proposes one safe next action through a guarded agent, surfaces safe memory guidance, and keeps the final decision in human review.
Architecture
The system is built as a monorepo with a Next.js product UI and modular packages for extraction, validation, RAG, agent workflow, memory, database, gateway, shared schemas, and evals. Postgres stores claim state, policy chunks, review tasks, memory records, AI call logs, and eval results. pgvector powers policy retrieval.
01
Claim intake
02
Extraction run
03
Deterministic validation
04
Policy RAG
05
Workflow memory
06
Guarded agent action
07
Human review
08
Trace and eval
Visual proof
These assets are pulled from the project repositories so the case studies show the actual product workflow, architecture diagrams, dashboards, and runtime proof instead of placeholder cards.
Architecture
Extraction, validation, RAG, memory, guarded agent action, human review, traces, and evals as one workflow.
Workflow proof
The reviewer starts from a claim source and creates a durable extraction run.
Workflow proof
Coverage reasoning is grounded in retrieved policy clauses and citation support.
Workflow proof
Memory provides safe workflow guidance without copying old claim facts.
Workflow proof
The model proposes one bounded action while backend guardrails control execution.
Product UI
The trace reconstructs the claim from intake through review and outcome.
Product workflow
Why it is not trivial
ClaimFlow AI proves that LLMs can assist high-stakes workflows only when surrounded by deterministic validation, policy evidence, bounded tools, memory safety, human review, evaluations, and observability.
Subsystem deep dive
The project exists to prove that applied AI can be designed as a reliable workflow rather than a set of disconnected demos. It answers how extraction, validation, RAG, an agent, memory, human review, evaluation, observability, and governance work together around one claim.
Coverage reasoning is grounded in policy evidence. The system loads a synthetic policy corpus, parses it into clause-level chunks, embeds them, stores them in Postgres with pgvector, retrieves relevant clauses, verifies citation support, and persists the coverage question.
The agent is not an open-ended autonomous loop. It is a controlled single-step tool-calling workflow. Backend state is loaded first, deterministic routing handles obvious states, and the model can propose exactly one registered tool call.
Memory stores reusable workflow lessons from trusted outcomes. It can guide caution or routing, but it is not claim evidence and cannot copy old claim facts into the current claim.
The AI gateway records model-backed calls with trace IDs, model metadata, prompt/schema versions, latency, token usage, cost, status, and failure classes. Evals test behavior across extraction, validation, review, RAG, agent guardrails, memory, and gateway observability.
Data model
Engineering decisions
What makes it more than a demo
Next improvements
Proof links
Proof links connect to the source repos, demos, architecture assets, screenshots, and engineering journals used to build this case study.
Final takeaway
ClaimFlow AI proves that LLMs can assist high-stakes workflows only when surrounded by deterministic validation, policy evidence, bounded tools, memory safety, human review, evaluations, and observability.