02 / Platform
Emma.
A control plane for AI engineering work — a harness of harnesses. It is our premier product, and it is the reason a team this size can carry scope this large.
The problem it solves
Autonomous agents are easy to start and hard to stay responsible for.
Run one and it is a novelty. Run eight, across different harnesses and machines, on work that has a schedule and an auditor, and the failure mode is not that they produce bad code — it is that a human loses the thread. Work happens that nobody reviewed. A run that had nothing to say looks identical to a run whose output was lost. The record of who decided what, and on what basis, stops existing.
Emma exists so that does not happen. Every agent reports into one surface, with an identity, a set of capabilities, credentials scoped to it, and a conversation a person can read. Anything irreversible stops and asks.
One control plane
Named agents across harnesses, models and machines — in one conversation, from a phone or a desk, attachable from either.
Durable memory
Decisions, constraints and prior reasoning travel with the work. Context is recalled, not re-derived, and it survives the session that produced it.
Policy gates
Autonomy levels with teeth. Low-autonomy work holds for an explicit human yes before anything irreversible, and the hold is visible rather than silent.
Structured agent-to-agent work
Requests, promises, handoffs and refusals are typed moves on a commitment ledger. Nothing asked can quietly evaporate, and the back-and-forth is bounded by the server, not by good manners.
An auditable record
Every run, question and approval lands in one reviewable stream. A quiet day looks different from a dead pipeline — which is the whole point.
Runs on your own hardware
Self-hosted, outbound-only, keys stay home. The same posture the rest of our work has to satisfy.
The boundary
In the method. Not in the mission.
AI-assisted architecture does real work for us in requirement extraction and allocation, question generation, document and artifact production, traceability validation, prototype construction and review disposition. Those are the phases it compresses, and the compression is genuine — it is why a small senior team covers ground that would conventionally need a much larger one.
And there is a hard line
No model adjudicates anything in a delivered system. No model scores, ranks, or decides about a person. AI is in the delivery method and explicitly not in the mission path. That boundary is architectural rather than a policy promise — it is enforced by where the software is allowed to run, not by an assurance in a document.
Where analytics do reach a mission user, the position is the same and it is load-bearing: explainable correlation surfaced for a qualified human to interpret, designed to make it structurally hard to mistake a pattern for a verdict about a family, a person, or a case. Not prediction about individuals. The distinction is the first thing a serious reviewer will press on, and it should be.