A memory layer for the org — and for its agents.
Ambito turns an organization’s communications into structured, persistent context — served to two consumers at once: the people doing the work (an ambient dashboard) and the AI agents assisting them (a permission-aware MCP server). The bet is that the durable value isn’t another model; it’s the context layer underneath them.
Problem
Org knowledge is scattered across threads, docs, and calls. People re-derive what the company already knows; agents start every task context-blind. Both need the same thing — a memory of what’s been decided and said, with permissions respected.
Role
Solo — this was my Maven AI PM capstone. I scoped it, locked the decisions, directed the specs, and built the backend. PM and builder in one seat.
Process
I scoped capability with the Overton Window method — scored 46 candidate capabilities, kept 27, cut 19 — then ran the Four-D Method, an AI PRD, and eval-driven error analysis (LLM-as-judge). Around it: an 11-competitor analysis, a three-model deployment architecture (SaaS / BYOK / VPC), and a pricing model benchmarked against Glean, Granola, and Copilot. The backend is built and tested — an end-to-end inference pipeline (received → classified → session-batched → extracted → escalated → persisted), 55 tests passing, on Postgres + pgvector with FastAPI ingestion; the FastMCP query server is in progress.
FIG 1.1the moat is one layer down from the frontier models: additive, not competitive. As models get better, a clean context layer gets more valuable, not less.
Selected for Product Faculty’s "Featured Five" — a 5-week documentary series following the build. Marketing site live at ambito.ai.
Pre-MVP: no live users, no metrics. The YC Summer 2026 application was submitted (May 2026) and is pending — not accepted. The $2.75M pre-seed deck is content-complete; no raise has been opened. The product itself is not built.