AgentLedger: Zero-Trust Accounting & Ledger Fleet

• 

  Detailed Model Context Protocol (MCP) tool execution steps (fetchLedgerAccounts, verifyTaxCompliance, postJournalEntry).

• 

  Dark-mode live dashboard overview with model badges, chart of accounts, and invoice ingest console.

• 

  High-tech multi-agent architecture diagram (Gemini 3.5 Pro/Flash, Gemma 2, MCP Tools, Cloud SQL).

• 

  Live ReAct trace of balanced AWS invoice verified and atomically POSTED to PostgreSQL.

• 

  Mathematical zero-trust rejection

• 

  unbalanced invoice logged to HITL queue. Interactive "Review & Post" Human-in-the-Loop override modal with debit/credit account allocation.

• 

  Gemma 2 2B IT Privacy Shield step redacting SSN, bank routing, and employee identity into tokens.

• 

  Visual infographic detailing the redaction and preservation rules.

💡 Inspiration

In enterprise finance and auditing, "AI hallucinations" aren't just inconvenient—they are catastrophic. If a financial LLM hallucinates an account balance or posts an unbalanced journal entry, the entire general ledger is corrupted.

We asked: How can we build an autonomous financial agent that organizations can trust with their balance sheet?

The answer was AgentLedger: a zero-trust multi-model fleet where mathematical constraints (abs(sum(debits) - sum(credits)) < 0.001), privacy redaction, and compliance gates are strictly enforced at the Model Context Protocol (MCP) tool layer.

⚙️ What It Does

1. Intelligent Multimodal Ingest (gemini-3.5-flash): Parses invoice scans, PDFs, and raw OCR text into strict, mathematically verified financial JSON schemas.
2. Privacy Shield (gemma-2-2b-it): Redacts bank account numbers, routing codes, employee identities, and SSNs into privacy tokens ([REDACTED_ACCOUNT], [REDACTED_SSN]) before reasoning.
3. Autonomous ReAct Audit (gemini-3.5-pro): Implements a multi-turn Reason + Act loop that queries the chart of accounts, audits tax compliance, maps debits and credits, and detects mismatches.
4. Zero-Trust Commitment & Rejection:
   • Balanced Entries: Atomically posted to Google Cloud SQL (PostgreSQL).
   • Unbalanced or Low-Confidence Entries (<85%): Rejected and queued for Human-in-the-Loop review.
5. Real-time SSE Dashboard & HITL Workflow: Streams live agent thoughts, tool calls, and observations with an interactive "Approve & Post" override card.

🛠️ How We Built It

• AI Models: gemini-3.5-pro (ReAct reasoning), gemini-3.5-flash (multimodal extraction), gemma-2-2b-it (PII redaction).
• SDKs & Frameworks: @google/genai (Google Gen AI TypeScript SDK), @modelcontextprotocol/sdk (Model Context Protocol), Antigravity CLI.
• Backend & Database: Node.js v20+, TypeScript, Express (Server-Sent Events), PostgreSQL with ACID transaction pooling.
• Google Cloud Infrastructure:
  • Google Cloud Run: Containerized microservice deployment.
  • Google Cloud SQL: Managed PostgreSQL 16 instance connected via native Cloud SQL Proxy Unix sockets.
  • Google Secret Manager: Secure injection of GEMINI_API_KEY and database credentials.
  • Google Cloud Build & Artifact Registry: Automated CI/CD container build pipeline.

🧗 Challenges We Ran Into

• Ensuring Zero Hallucinations: Prompt instructions alone are not enough for double-entry bookkeeping. We solved this by enforcing zero-trust checks at the tool execution level inside src/mcp/ledgerTools.ts, making it impossible for the agent to commit an unbalanced transaction.
• PII Redaction without Corrupting Financial Numbers: Distinguishing between account numbers (which must be redacted) and financial totals/quantities (which must be preserved). We paired Gemma 2 2B IT with strict preservation rules.

🏅 Accomplishments That We're Proud Of

• 100% Pass Rate on Zero-Trust Audit Suite: 9/9 automated checks verifying mathematical balance, low-confidence safety thresholds, and PII protection.
• Multi-Model Google Fleet: Seamless pipeline orchestrating Gemma 2 2B IT $\rightarrow$ Gemini 3.5 Flash $\rightarrow$ Gemini 3.5 Pro.
• Zero-Downtime Cloud Run Deployment: Single-command script (./deploy.sh) connecting Cloud Run to Cloud SQL with Secret Manager integration.

📚 What We Learned

Enterprise-grade agentic workflows require a defense-in-depth architecture: combining fast models for extraction, specialized edge models for privacy, reasoning models for orchestration, and hard deterministic boundaries at the tool layer.

🔮 What's Next for AgentLedger

• ERP connectors for NetSuite, SAP, and QuickBooks via MCP sidecars.
• Automated multi-currency foreign exchange (FX) variance revaluation.
• Autonomous vendor discrepancy communication agent.

Built With

• accounting
• agentic-ai
• cloud-build
• express.js
• fintech
• gemini-3.5-flash
• gemini-3.5-pro
• gemini-api
• gemma-2
• google-cloud-run
• google-cloud-sql
• google-secret-manager
• mcp
• model-context-protocol
• node.js
• postgresql
• typescript