Lumina — B2B AI Reconciliation Agent

Autonomous B2B financial reconciliation powered by Google ADK multi-agent pipeline, Gemini 3.5 Flash & MongoDB Atlas MCP. Cuts 3-day cycles to seconds with Human-in-the-Loop approval.

Comment

The Problem

Every month, finance teams across the world lose 3–5 business days manually reconciling inter-company account balances. A typical B2B reconciliation cycle looks like this: export ledgers to Excel, email your counterparty, wait for their response, spot-check line by line, draft resolution emails manually — and repeat for every discrepancy. Human error rates exceed 12%, and there's no audit trail.

This is a $4.7B+ annual problem that nobody has solved with AI agents — until now.

What We Built

Lumina is a production-ready B2B financial reconciliation platform powered by a Google ADK 2.0 multi-agent pipeline, Gemini 3.5 Flash, and MongoDB Atlas as the core data layer — connected to agents via a custom MCP (Model Context Protocol) server over real HTTP/SSE transport.

The entire reconciliation cycle — from ledger import to discrepancy detection to email dispatch — runs in under 8 seconds.

Architecture

ERP / Excel / CSV ↓ FastAPI (Google Cloud Run) + JWT Auth ↓ lumina_orchestrator (Gemini 3.5 Flash) ├── reconciliation_agent → MCP find() → MongoDB ledgers ├── analysis_agent → MCP aggregate() → classify discrepancies └── communication_agent → MCP insert() → draft resolution emails ↓ Human-in-the-Loop approval gate ↓ SMTP email dispatch → case closed

MongoDB is not just storage — it's the agent's superpowers. Every agent decision is a structured MCP tool call (find, aggregate, insert_one, update_one, vector_search) against MongoDB Atlas collections. The agents never see raw data dumps — they query exactly what they need, mid-reasoning.

Key Features

  • 🤖 Multi-agent orchestration — 4 specialized agents with defined roles under a root orchestrator
  • 🔗 Real MCP server — HTTP/SSE transport at /mcp/sse, full JSON-RPC handshake, tool discovery
  • 🔍 Atlas Vector Search — semantic similarity matching via text-embedding-004 for fuzzy record deduplication
  • 🌐 B2B Portal system — counterparties receive a secure, single-use magic link, review your ledger entries, and respond (Agree / Dispute / Let AI Compare)
  • 👤 Human-in-the-Loop — every AI-drafted email requires explicit approval before dispatch (critical for financial contexts)
  • 🏭 Local ERP Agent — downloadable agent for SAP, Logo, Mikro — syncs on-premise data to MongoDB Atlas automatically
  • 📊 Real-time dashboard — discrepancy feed, trend heatmap, agent execution timeline
  • 💱 Multi-currency support — TRY, USD, EUR, AUD, NOK, CNY with sign-aware matching

How We Built It

Backend: FastAPI on Google Cloud Run, Motor async MongoDB driver, Google ADK 2.0 SDK, custom MCP ASGI app mounted at /mcp/sse.

Frontend: Next.js 14 on Vercel, TanStack Query, D3.js world map, real-time AgentIsland execution panel.

AI: Gemini 3.5 Flash via Google Cloud Agent Builder SDK for all agent reasoning; text-embedding-004 for vector embeddings.

Data: MongoDB Atlas with 7 collections (ledgers, companies, discrepancies, master_balances, reconciliation_sessions, agent_runs, file_objects), Atlas Vector Search index on ledger descriptions.

Challenges We Faced

1. MCP over HTTP/SSE in production — Most examples use subprocess transport. We implemented a real HTTP/SSE MCP server as an ASGI app inside FastAPI, with proper session lifecycle and JSON-RPC message routing. Getting this to survive Google Cloud Run cold starts required OpenSSL SECLEVEL patches and Motor async connection pooling.

2. MongoDB Atlas TLS on Cloud Run — OpenSSL 3.0 cipher incompatibilities caused silent SSL failures. Fixed with SECLEVEL=1 patch, certifi CA bundle, and python:3.12-bookworm base image.

3. Agents hallucinating financial figures — Early versions would invent amounts when context was sparse. Solved by routing all data through MCP tool calls — agents never receive raw data in prompts, only structured tool results from MongoDB.

4. Human-in-the-Loop UX — Getting the approval gate to feel natural (not like an obstacle) required significant UI iteration. The final design shows side-by-side amount diffs, AI fix suggestions, and the auto-drafted email — all in one modal — so approving takes one click with full context.

What We Learned

  • MCP as a data bridge is genuinely transformative for agent accuracy. Structured tool calls eliminate hallucination on numerical data better than any prompt engineering.
  • Google ADK's agent hierarchy (root orchestrator → sub-agents) maps perfectly to real enterprise workflows where different "departments" (matching, analysis, communication) have distinct responsibilities.
  • MongoDB Atlas is an ideal agent data layer — the same database serves real-time queries, vector search, aggregation pipelines, and audit logs, all through one connection.
  • Financial AI requires Human-in-the-Loop — not as a safety measure, but as a product feature. Accountants trust the AI more when they have final say.

Impact

Metric Manual Process Lumina
Time per reconciliation cycle 3–5 business days < 8 seconds
Human error rate > 12% < 1%
Audit trail Spreadsheets Full MongoDB history
Email drafting Manual, inconsistent AI-generated, professional

Lumina targets the 400M+ SMEs globally that run reconciliation manually — a segment too large for enterprise ERPs, too complex for spreadsheets.

Built With

  • d3.js
  • fastapi
  • gemini-3.5-flash
  • googlecloudadk
  • googlecloudrun
  • mcp
  • mongodb-atlas
  • mongodb-vector-search
  • motor
  • next.js
  • python
  • tanstack-query
  • typescript
  • vercel
Share this project:

Updates