AI Agent Reliability Cockpit

An interactive demo that proves AI agents in enterprise audit can be controlled, transparent, and measurably improvable — without model retraining or black-box infrastructure.

Built for C-Suite, Audit Leaders, and Compliance Officers.

---

What it demonstrates

Step	Page	What the audience sees
1	Overview	Four pillars of reliable AI + the improvement loop
2	Run Agent	Real audit case — agent returns Medium Risk at 68% confidence
3	Trace Viewer	Every step the agent took, with tool names, inputs, outputs, confidence
4	Evidence Map	Every claim grounded to a source — missing evidence flagged explicitly
5	Human Review	Expert overrides to High Risk, proposes a rule
6	Learning Log	Rule captured instantly — no retraining
7	Agent Rerun	Same case → High Risk at 91% confidence, rule cited
8	Evaluation Dashboard	Before/after: Accuracy +28pp, Time −66%
9	Business Impact + Architecture	₹19L+ savings per 1,000 cases, operating model

The demo is fully stateful — running the agent, submitting feedback, and rerunning all flow through shared React context. "Reset Demo" restores the initial state for repeat runs.

---

Stack

• Frontend: React 18 + Vite + TypeScript + Tailwind CSS v4
• Charts: Recharts
• Routing: Wouter
• Monorepo: pnpm workspaces (Node.js 24)
• Backend: Express 5 (API server — not used by the demo itself, included as workspace scaffolding)

---

Project structure

.
├── artifacts/
│   ├── cockpit/          # The demo app (React + Vite)
│   │   └── src/
│   │       ├── pages/    # One file per demo page (9 pages)
│   │       ├── context/  # DemoContext — shared stateful demo flow
│   │       └── components/  # Layout, sidebar
│   ├── api-server/       # Express API server (workspace scaffolding)
│   └── mockup-sandbox/   # Component preview server (workspace tooling)
├── lib/                  # Shared TypeScript libraries
├── scripts/              # Utility scripts
├── pnpm-workspace.yaml   # Workspace config + catalog pins
├── tsconfig.base.json    # Shared strict TS config
└── tsconfig.json         # Solution file (libs only)

---

Getting started

Prerequisites

• Node.js 24+
• pnpm 9+

Install

pnpm install

Run the demo app

pnpm --filter @workspace/cockpit run dev

Open http://localhost:5173 in your browser.

Run the API server (optional)

pnpm --filter @workspace/api-server run dev

Typecheck everything

pnpm run typecheck

---

Demo flow

The demo uses a shared DemoContext with four states:

initial → ran_initial → feedback_submitted → reran

1. Run Agent page — click "Run Initial Agent" → transitions to ran_initial
2. Human Review page — click "Submit Feedback" → transitions to feedback_submitted
3. Run Agent page — "Rerun After Feedback" button unlocks → click → transitions to reran
4. All other pages (Trace Viewer, Evidence Map, Learning Log, Evaluation, etc.) react to the current state

Click Reset Demo on the Run Agent page to return to initial.

---

Audience variations

Audience	Focus
C-Suite / CFO	Business Impact → Evaluation Dashboard → skip Architecture
Audit Leaders	Evidence Map → Learning Log → Audit Trail metric
Compliance Officers	Trace Viewer → Architecture governance layers
Technical	Trace Viewer (tool calls, confidence scores, IF-THEN rule syntax)
5-minute version	Run Agent → Trace Viewer → Human Review → Rerun → Metrics

---

Key talking points

1. "Traceability is the foundation." — Trace Viewer shows what it did. Evidence Map shows where it got information.
2. "Explainability is proof." — We don't just say High Risk. We explain why.
3. "Review is control." — The human decides. The agent suggests.
4. "Learning is improvement." — Each correction becomes a rule for future cases.
5. "Metrics are accountability." — We measure. We improve. We report.

---

License

Internal demo — not for public distribution.