SentinelFlow

Architecture

SentinelFlow

SentinelFlow is a governed pre-ledger financial control plane that stops corrupted ACH and payment files before ledger impact, then uses a bounded Gemini 3.5 + Google ADK agent fleet to investigate, explain, and propose safe remediation without giving AI financial authority.

The problem

A one-cent imbalance, duplicate record, malformed trailer, or broken hash total in a high-value payment file can trigger reversals, reconciliation work, SLA breaches, and regulatory risk. Traditional validators flag errors but rarely preserve a complete evidence chain or coordinate a trustworthy investigation.

The solution

Deterministic Go services ingest and validate each file, quarantine failures, build tamper-evident evidence, score policy risk, and enforce independent verification plus dual-control release. A governed Google ADK specialist topology uses Gemini 3.5 Flash through a hardened model boundary to explain incidents and propose next actions. Agent output is advisory; deterministic controls and humans retain authority.

The demo opens with a memorable proof: a synthetic $4.7 million ACH file is quarantined because of a $0.01 imbalance, with seven blocking findings preserved as evidence.

Key features

Deterministic pre-ledger validation and quarantine
Tamper-evident evidence bundles and canonical hashes
Governed Google ADK specialist fleet
Gemini 3.5 Flash through Google GenAI SDK and Vertex AI
Model Armor pre/post screening
Independent verification and dual-control release
ACH return intelligence
Cloud Run UI and gateway with Vertex AI Agent Runtime

Architecture

The React/TypeScript UI calls a Go gateway on Cloud Run. The gateway owns ingestion, validation, evidence, policy, verification, and release state. It delegates bounded advisory work to the Python Google ADK tier, which invokes Gemini 3.5 Flash through the GuardedModelBoundary and Model Armor. Human review and deterministic verification gate every financial state transition.

Testing

The README contains reproducible setup and test instructions. Synthetic fixtures are under demo-data/. Run python scripts/hackathon_demo_proof.py for four live checks: quarantine, Gemini 3.5 Flash, Vertex AI Agent Runtime, and Model Armor.

Built with Codex

Codex helped audit trust boundaries, harden provider and guardrail paths, add tests, inspect live Google Cloud deployments, verify evidence, improve reproducible setup instructions, and assemble the demo and submission.

Known limitations

The public demo uses synthetic financial data. The short video proves the managed IncidentCommanderAgent and governed boundaries; the broader specialist topology is implemented and tested but not every specialist is exercised live. Agent memory is advisory, never authoritative evidence.