Deal Desk OS — an AI co-pilot for powersports dealership F&I teams.
Drop in raw deal notes, get back a structured deal summary, suggested F&I add-ons, compliance flags, and a ready-to-send customer follow-up SMS — streamed token-by-token in seconds.
▶ Live demo · No login, no API key — runs entirely in your browser.
▲ RevAssist Pro on Vercel · Next.js fullstack deployment with signed demo sessions, OIDC/JWKS managed-auth adapter, streaming API, Neon persistence, Upstash rate limits, eval-gated mock AI mode, and remote smoke coverage.
- Architecture — product boundary, current frontend architecture, production backend target, API contract, data model, safety, and eval strategy.
- RevAssist Pro Case Study — product problem, architecture decisions, tradeoffs, security, evals, scalability, and interview talking points.
- RevAssist Pro — Next.js fullstack app with streaming API, signed demo sessions, OIDC/JWKS managed-auth adapter, Neon/Postgres persistence, schema validation, Upstash-backed rate limits, audit events, history, tests, and live AI mode.
- RevAssist Pro Eval Baseline — deterministic regression report for routing, schema validity, add-on relevance, compliance coverage, and SMS quality.
- RevAssist Pro Live Eval Snapshot — provider-backed eval snapshot workflow for model/prompt launch gates.
- RevAssist Pro Deployment — Vercel, Neon, Upstash, env vars, observability, launch checklist, and rollback runbook.
- RevAssist Pro Authentication — managed-auth adapter and rollout plan for provider sessions, dealership membership, middleware, and tenant-safe claims.
- RevAssist Pro Release Checklist — quality gates, public sharing checklist, rollback, and next release candidates.
- RevAssist Pro v0.1.0 Release Notes — launch summary, engineering highlights, evidence links, and known limits.
- Changelog — release history and known limitations.
- Security Policy — supported surfaces, reporting path, and security posture.
Powersports F&I managers spend hours every week retyping the same customer info into lender portals, hand-writing follow-up texts, and remembering which add-ons make sense for which customer profile. RevAssist collapses that work into one structured AI response.
It's a focused exploration of what AI-augmented internal tooling looks like inside the modern dealership stack — built for the people who actually close deals, not for the marketing team.
Built for the brands the dealer actually carries: Yamaha, Polaris, Can-Am, Sea-Doo, Honda, Kawasaki, Harley-Davidson.
Given a deal description like:
"Customer wants a 2024 Yamaha YZF-R1, $2k down, 60-month term, 720 FICO. First-time sportbike buyer, no trade-in."
RevAssist returns a schema-locked JSON response containing:
- Deal summary — 2–3 sentence plain-English recap
- Suggested add-ons — 3 F&I products tailored to the customer profile, with rationale and price ranges
- Compliance flags — items to verify, color-coded by severity (info / warn / block)
- Customer follow-up SMS — a warm, professional 1–2 sentence text ready to copy and send
All four sections render progressively as the response streams in.
This build runs in mocked mode — the streaming response is simulated client-side so the demo runs without an API key in the browser. In production, the LLM call would live behind your own backend with auth, rate limiting, audit logging, and proper key management. The frontend patterns (partial-parse rendering, token-by-token UX, schema-locked output) are the part being demonstrated.
cd app
npm install
npm run devThen open http://localhost:5173. No API key required.
For the production validation suite:
npm run lint
npm run test
npm run build
npx playwright install chromium
npm run smoke- React 19 with hooks
- Vite 8 for the dev server and bundle
- Tailwind CSS v4 via the official Vite plugin
- Lucide React for iconography
- Barlow Condensed + Inter + JetBrains Mono type stack
- Vitest coverage for deal routing, schema validation, partial JSON parsing, and export formatting
- Playwright production smoke tests for the full deal workflow on desktop and mobile
- Streaming LLM backend for structured JSON generation (production)
- Token-streaming UX with partial-JSON parsing as content arrives
- Isolated deal engine: sample deals, mock response selection, schema validation, partial parsing, and copy/export formatting live in
src/lib/dealEngine.jsso the product logic can be tested independently of React. - Schema-locked output: the system prompt forces the model to return strict JSON. The frontend parses the stream incrementally and renders each section as soon as it's structurally valid.
- Optimistic partial render: as tokens arrive, a parse attempt runs on every chunk. The first valid parse swaps the raw stream view for the structured render.
- Production backend path: docs/ARCHITECTURE.md documents the authenticated streaming API, data model, rate limits, audit logs, and evals needed to turn the browser demo into a SaaS workflow.
- Case-study narrative: docs/case-studies/revassist-pro.md explains the engineering decisions, tradeoffs, and interview story behind the Pro app.
- Eval baseline: pro/docs/EVAL_BASELINE.md turns the regression suite into a readable quality report for reviewers.
- Authentication: pro/docs/AUTHENTICATION.md explains how signed demo sessions and managed provider JWTs map into one tenant-safe claim contract.
- Release readiness: pro/docs/RELEASE_CHECKLIST.md documents launch gates, rollback, and next release candidates.
- Latency surfaced: response time and token count are exposed in the UI for transparency.
- Three sample deals (sportbike / UTV / PWC) included for fast testing.
The main branch auto-deploys to GitHub Pages via .github/workflows/deploy.yml. The workflow installs from lockfile, runs a production audit, lints, runs unit tests, builds the Vite bundle, runs Chromium smoke tests, then deploys the artifact. To enable:
- Push to
main. - In the repo: Settings → Pages → Source: GitHub Actions.
- The next push triggers a build; the URL appears in the workflow run.
For root-hosted deploys (Vercel, Netlify), set VITE_BASE_PATH=/ in the build environment.
- Connect a hosted sign-in UI and provider org sync to the managed-auth adapter.
- Add provider-backed live eval snapshots before enabling live AI by default.
- Add DMS / credit-bureau integrations to pre-fill from a real lead.
- Add voice input for the deal desk.
Built by Zoriah Cocio — info@zoriahcocio.com
