FinAgent — B2B KYC & Due Diligence AI

• 

  Knowledge Graph Integration, AI Agent with Knowdledge Graph Integration

• 

  INVESTIGATOR AGENT

• 

  WORKFLOW AI AGENT

The payment decision is inside the agent's reasoning loop, not in the UI layer. The KYCAgentState TypedDict carries payment_status, investigation_depth, doku_link, cypher_query, graph_context, and answer across nodes.

Stack

• FastAPI + Jinja2 — REST API + server-rendered SPA frontend
• LangGraph — agentic workflow orchestration with conditional edges
• GPT-4o — entity extraction, Cypher generation, report synthesis
• Neo4j AuraDB — cloud knowledge graph with Cypher queries
• DOKU Checkout v1 API — real sandbox payments with HMAC-SHA256 signing
• vis.js Network — interactive graph visualisation
• rapidfuzz + jellyfish — ICIJ-inspired entity deduplication

Challenges we ran into

1. Cypher Hallucination — GPT-4o generated invalid Cypher syntax (type(var)-[:REL]->() in RETURN, deprecated size() for Neo4j 5). Built a _sanitize_cypher() sanitiser and extended the system prompt with explicit anti-hallucination rules and Cypher examples.

2. DOKU Webhook Reachability — Microsoft Dev Tunnels requires authentication so DOKU's servers couldn't POST to our local webhook. Built a dual confirmation system: server-side webhook for production + browser redirect (/payment-success) as reliable fallback.

3. Session Loss on Hot-Reload — In-memory _sessions dict wiped on every code change. Solved with localStorage on the frontend + file-backed sessions.json on the backend with session resurrection logic when the server restarts mid-investigation.

4. Chat History Across Payment Redirect — DOKU redirects the browser away and back, wiping JS memory. Persisted _chatHistory (last 50 messages) to localStorage and restored on page load with _resumeAfterPayment flag.

Accomplishments that we're proud of

• Payment as a first-class agent node — not a UI trick, but a real conditional edge in LangGraph that the agent evaluates on every invocation
• Real DOKU sandbox integration working end-to-end with HMAC-SHA256 signature generation, live checkout pages, and payment confirmation flow
• Agent Trace accordion in chat — every LangGraph node's output (query plan, Cypher query, row count, synthesis) is visible and expandable for full transparency
• Production-grade GraphRAG — multi-hop Cypher traversal catches shell company chains that vector similarity search would miss entirely
• Full UI with interactive knowledge graph, drag-to-resize panels, export to HTML report, and persistent chat history — all in under 12 hours

What we learned

• LangGraph conditional edges are the cleanest way to implement business logic gates inside an AI agent — far more maintainable than if/else chains in a single prompt
• GraphRAG > VectorRAG for structured relationship queries — knowing that Company A directly BORROWS_FROM Company B is more precise than semantic similarity
• Embedding payment confirmation at the agent reasoning level opens entirely new monetisation models for AI SaaS — every capability tier can be a node
• Neo4j AuraDB's Cypher dialect requires careful prompt engineering to avoid Neo4j 4 vs 5 compatibility issues

What's next for FinAgent — B2B KYC & Due Diligence AI

• DOKU production keys — switch from sandbox to live payments for real B2B clients
• Multi-document investigation — cross-reference entities across multiple uploaded documents simultaneously
  
• Automated red flag alerts — periodic re-analysis that notifies users when new connections appear in the graph
• AML scoring model — fine-tuned risk scoring based on transaction patterns, jurisdiction risk, and network centrality metrics
• API access tier — expose /api/investigate as a paid B2B API for banks and compliance teams to integrate into their existing workflows
• Graph diff view — show what changed in the knowledge graph between two investigation runs

Built With

• doku
• fastapi
• groq
• knowledgegraph
• langchain
• langgraph
• neo4j
• ngrok
• openai
• openclaw
• python
• sumopod