SecureBank — Human-in-the-Loop AI Support Agent (prototype)

An AI support agent for a financial services mobile app. A customer reports a problem (by chat or a simulated voice call); the agent reads it, reproduces it in an isolated sandbox of a bundled mock banking app, proposes a fix, and routes that fix to a human approval gate. Nothing is applied to the app until a human approves. All customer data is PII-redacted and every step is logged for transparency.

v1 is fully offline and deterministic: the "LLM" and "speech-to-text" are mocked so the whole flow runs with no API keys and is easy to test. See Swapping in real models below for where Claude + a real transcription service drop in.

Screenshots

Customer chat — a complaint is sent, and the agent classifies, reproduces the bug in the sandbox, and proposes a fix (right panel shows the audit trail):

Voice call — the same pipeline driven by a simulated call with live transcript:

Approval console — the human gate where proposed fixes are approved (auto-verified in sandbox) or rejected:

The workflow

Customer (chat OR voice)
        │  raw text  ─────────────►  🔒 PII redaction (SSN, card, email, phone, acct#)
        ▼
   1. CLASSIFY  ──► no match ──► ASK ONE CLARIFYING QUESTION ──► still unclear (x2) ──► ESCALATE to human
        │ matched
        ▼
   2. REPRODUCE in isolated sandbox (fresh state each run)
        │ bug observed                 │ no failure
        ▼                              └──► clarify / ESCALATE
   3. PROPOSE FIX (code diff + user steps + risk + repro evidence)
        ▼
   4. ⛔ HUMAN APPROVAL GATE ◄── the machine stops here, always
        │ approve                      │ reject
        ▼                              └──► ESCALATE
   5. APPLY in sandbox ──► VERIFY (re-run fixed) ──► pass? RESOLVED : ESCALATE

Voice uses the exact same state machine — it just transcribes the call first (streaming interim partials to mimic live captioning), then feeds the final transcript into step 1.

Guarantees built in

• Human gate is absolute. Code stops at awaiting_approval; apply/verify only run inside decide() after an approval. A test asserts no fix is applied pre-approval.
• Data security. Raw customer text/transcripts never leave the server — the API only ever returns redacted views (toView() strips raw* fields).
• Transparency. Each ticket carries a human-readable timeline and a structured audit log, both shown in the dashboard.

The bundled mock app (what gets "reproduced")

Four real, seeded bugs in server/src/bank.ts, each with a buggy + fixed code path so reproduction and post-fix verification are genuine, not narrated:

Bug	Symptom
transfer-overdraft	Dollar-rounding lets a transfer overdraw the account
login-early-lockout	Lock counter never resets on success → locks too early
balance-rounding	Balance rounds to whole dollars, hiding cents
statement-empty-crash	Statement export crashes on an account with no transactions

Run it

cd support-agent
npm install          # installs server + web (npm workspaces)
npm run dev          # starts API (:4000) and dashboard (:5173)

Open http://localhost:5173.

Test it thoroughly

npm test             # vitest: bug repro/fix, PII redaction, the human gate, clarify→escalate

Try in the dashboard

1. Customer chat tab → click a sample complaint chip → Send to support → watch the agent classify → reproduce → propose on the right.
2. Voice call tab → pick a spoken complaint → consent → Call in → watch the live transcript, then the same flow.
3. Approval console tab → Approve & apply (auto-verifies) or Reject → escalate.
4. Try the vague complaint to see the clarify→escalate path.

Swapping in real models (production seam)

Everything model-shaped is isolated in server/src/agent.ts:

• classify(text) → replace keyword match with a Claude call (return the same Classification shape).
• buildPartials() / voice transcript → replace with a streaming speech-to-text service; keep emitting partials into ticket.transcript.
• The reproduction target (server/src/bank.ts) → replace the bundled app with an adapter that drives your real staging environment / test harness.

The orchestrator, human gate, redaction, audit trail, and dashboard stay unchanged.

Layout

server/src
  bank.ts          mock app + 4 seeded bugs (buggy & fixed paths)
  agent.ts         mocked LLM classify + proposal + voice presets  ← model seam
  security.ts      PII redaction
  orchestrator.ts  the workflow state machine + HUMAN GATE
  store.ts         in-memory tickets + redacted API projection
  routes.ts        chat / voice / approvals HTTP API
  workflow.test.ts vitest coverage
web/src
  App.tsx          tabs, ticket list, live polling
  components/      CustomerChat, VoiceSim, ApprovalConsole, TicketDetail