Erick Manrique
01Signal

Ambito

Deployed prototype · staging · no external users

A permission-aware context layer that turns an organization’s communications into structured, persistent memory — served to both the people and the agents doing the work.

  • Postgres + pgvector
  • Permission-aware MCP server
  • Haiku classifier
  • Sonnet extractor
  • Opus judge
  • Gmail · GCal · Zoom
Problem

Work fragments across four to six tools, and context evaporates twice over: decisions and commitments never make it out of the communications they happened in — the inference gap — and the canonical knowledge that does survive is unreachable by the AI agents that need it — the canonical gap. People re-derive what the org already decided; agents answer confidently without the context to be right.

02System

How it works

Communications enter through connectors, get classified and extracted by a three-model pipeline, resolve into a hierarchy, and land in one store that serves both a dashboard and an MCP server. One path, no side doors.

FIG 2.1 — data flow, ingest to serve

Band AIngest

  1. ConnectorsGmail · GCal · Zoom
  2. Ingestion workersqueue · normalize

Band BUnderstand

  1. ClassifierHaiku
  2. ExtractorSonnet
  3. JudgeOpus
  4. Hierarchy resolverentity · thread · org

Band CServe

  1. Postgres + pgvectorstructured · persistent
  2. Dashboardfor people
  3. MCP serverfor agents · permission-aware
03Process

How it was built

Built by the same two-role system the build record describes: a planning session that never writes code, a coding agent that never makes product calls, and two files between them.

FIG 3.1 — handoff timeline
  1. Step 01Planning session

    Reads the record, proposes scope, never touches code.

  2. Step 02Handoff brief

    The one file planning writes to. Code reads only this.

  3. Step 03Coding agent

    Implements against the spec, writes the failing test first.

  4. Step 04CI gate

    Checks required; PRs auto-merge on green. A red suite means no merge.

  5. Step 05Return channel

    Code reports back; the decision is appended, immutable from there.

FIG 3.2 — decision log, excerpts
IDDateDecisionVerdict
DEC-0772026-05-17Adopted a five-gate viability framework after a CONDITIONAL-NO verdict.Locked
DEC-109 → DEC-1152026-05-31Direct branch-merge-to-main reversed — a PR-based fleet workflow behind required CI.Reversed · locked
DEC-1692026-06-13Kill-switches and alerting relabelled PLANNED rather than delivered — the deliverable language was corrected in the spec.Locked

225 locked decisions on record · showing threeAppend-only · no edits, no deletes

04Evidence

What is on the record

One number carries the argument: the suite is large enough that the three-model pipeline can be changed without guessing at what broke. The rest is stated, not counted.

FIG 4.11,145

backend tests passing, plus 120 end-to-end

FIG 4.2Permission-aware MCP server

Agents read only what the requesting person can read. The permission model is the product, not a setting.

FIG 4.3Three-model pipeline

Haiku classifies, Sonnet extracts, Opus judges — cheap models filter, the expensive one only rules.

FIG 4.4225 locked decisions

Every material call is in the log with the spec it changed.

FIG 4.5Solo build, ~3.5 months

One person, agent-assisted throughout, under the operating system on the build record.

FIG 4.6Staging only

Production promotion has not been run.

Linksambito.ai
05Status & caveats

Where it stands

Deployed prototype · staging · no external users

Built and deployed to staging, prod-verified on a seeded tenant in June 2026 — and deliberately pre-launch: no external users, no production promotion, no revenue. The $2.75M pre-seed deck is content-complete; no raise has been opened.

Not claimed

  • Users, active or waitlisted
  • Revenue or pipeline
  • Promotion to production
  • Third-party validation