REX TrustDesk —Auditable AI for Financial Access © FoundLab

AI-native trust desk for financial access. Gemini agents run intake, risk and evidence live on Google Cloud — every decision leaves a tamper-evident, auditable receipt.

Comment

Inspiration

Financial-access decisions — onboarding, credit pre-screening, compliance intake — are slow, costly, and opaque. Small operations can't afford a compliance team, and when they bolt AI onto the problem, decisions turn into a black box nobody can audit. We wanted the opposite: let AI operate the workflow end-to-end, but make every decision verifiable with a receipt — so trust is proven, not assumed.

What it does

REX TrustDesk is an AI-native trust desk for financial-access workflows. A case flows through:

case intake → intake agent → risk agent → deterministic policy gate → evidence agent → DecisionReceipt

  • Gemini agents extract the case facts, classify risk, and draft the customer follow-up.
  • A deterministic policy gate — never the model — makes the final, material routing decision.
  • Every agent step logs an AgentEvent, and each case ends in a hashed, tamper-evident DecisionReceipt (input_hash, agent_trace_hash, receipt_hash, the exact policy rules evaluated, and the human-review status).

It runs live in production on Google Cloud Run, with the AI executing on Vertex AI (Gemini 2.5 Flash). The /health endpoint reports model_mode: "gemini" when the AI path is live, and a self-contained dashboard shows the agent trace, policy table, DecisionID, and receipt hashes for every case.

How we built it

  • A TypeScript pnpm monorepo: a receipt engine (canonical hashing), a deterministic policy-gate package, an agent runtime with a pluggable ModelAdapter, a shared processCase() orchestrator, and a JSONL evidence store.
  • The ModelAdapter has two Google backends — Vertex AI (keyless, via the Cloud Run service account) and the Google AI API — plus a deterministic fallback. Model output is validated before it can touch policy, using Gemini's native responseSchema plus our own validators.
  • Fastify API + a zero-build dashboard, packaged in a single container and deployed to Cloud Run.
  • Keyless CI/CD: GitHub Actions + Workload Identity Federation builds from source with Cloud Build and deploys on every push to main, then smoke-tests /health, /cases, and /evidence/summary.

Challenges we ran into

  • Safety of AI decisions. An LLM must not be able to approve anyone. We made the policy gate the sole decision authority and designed the model path to fail closed: on any model failure or invalid output, it falls back to deterministic logic and escalates risk to manual review — a broken model call can never produce an approval.
  • Keyless auth under strict org policy. The Google Cloud org blocks service-account key creation. We moved the whole pipeline to Workload Identity Federation, so there are no secrets anywhere — not in the repo, not in GitHub.
  • Not leaking raw data. Only derived, hashed artifacts ever leave the trust boundary; raw payloads are never logged or persisted.

What we learned

Making AI auditable is a design constraint, not an afterthought. The strongest pattern was "the model advises, the policy gate decides," with a hashed receipt as the contract between them — it keeps the system AI-native and trustworthy at the same time.

What's next

Onboarding paid pilots with small financial operators, durable evidence storage (Firestore/BigQuery), and richer risk policies — while keeping the auditable spine unchanged.

Synthetic demo data throughout — no real customer documents or PII. REX TrustDesk produces auditable candidate decisions with a human-review path; it does not auto-approve credit and makes no guarantee of regulatory compliance.

Built With

Share this project:

Updates