Deal Room

Live

🏆 Best B2B Sales AI — buyer portal with agents + evals

Started: June 18, 2026 · ~3 days of intensive build over the capstone window

AI:

Claude Code

Azure AI Foundry

Stack:

Next.js 16

React 19

TypeScript

Azure AI Foundry

Supabase

Vercel

White-label Digital Sales Room for B2B deals — a multi-tenant, login-gated buyer portal where buying groups see status, next steps, docs, and quotes, and sellers see engagement analytics. Two role-scoped Azure AI Foundry agents (buyer Q&A + AE analytics), both RAG-grounded, with a trust gate that requires human approval before AI drafts reach buyers. Built as the final capstone for the Maven AIPM Course by Mahesh Yadav — and it took home the Best B2B Sales AI award. Every showcase pillar has receipts: agents (live chat), evaluations (green leak-gate scorecard — 0% leak, 100% retrieval precision, 100% answer match), guardrails (RLS tenant isolation + access-scoped retrieval + Azure Content Safety + trust-gated authoring).

// demo video

6:38

// screens

1 / 2

// highlights

Won 🏆 Best B2B Sales AI at the Maven AIPM capstone showcase
Two role-scoped Azure AI Foundry agents — buyer deal-Q&A (RAG over buyer-visible docs only) and AE engagement/analytics — called via the OpenAI-compatible Responses API with AAD bearer auth
Guardrails at three layers: Supabase RLS tenant isolation, access-scoped RAG retrieval (seller-only chunks never surface to buyers), and trust-gated authoring — a human approves every AI draft before a buyer sees it
Self-contained evaluations harness exercises the real buyer/AE retrieval + agent paths against live Azure — hard ship-gate on 0% leak rate, with a positive control against seeded seller-only canaries so an empty index can't fake a pass
Green ship-gate scorecard: leak rate 0%, retrieval precision 100%, answer match 100% — receipts, not vibes
Two-truths guardrail demo: the same question runs through the buyer and AE stacks side-by-side, live — the buyer answer never contains the seller-only chunk
Beachhead vertical: aerospace/defense and mission-critical hardware — the regulated-technical wedge where trust-gated agentic authoring earns its keep

// takeaways

Evaluations are what separate 'looks safe' from 'is safe.' A hard ship-gate on leak rate — exit non-zero on any leak — plus a positive control against seeded canaries turned the guardrail claim from a story into a receipt.
Azure AI Foundry agents are referenced by NAME via the OpenAI-compatible Responses API — versionless endpoint, AAD bearer, no asst_ ids. Every instinct from the older Azure OpenAI shape 400s. Documented so the next builder skips that debugging arc.
Trust-gated authoring — human approval before AI content reaches the buyer — is slower than full autonomy, and that's the point. In regulated verticals it's the only shape the customer will trust.
Spec-driven building with parallel subagents works: 29 workstreams shipped by writing the spec first, fanning out disjoint agents, and always running the integrated build before committing — tsc alone doesn't catch RSC boundary breaks.

Live Demo ↗