ai158z · the stack
Each layer feeds the next.
ai158z is not a collection of experiments. It's one stack: a harness that makes models work, a workforce that sells that work, a refinery that compresses what the workforce learns, and a research line aiming those compressed minds at the hardware of the next decade.
01
shipped · in production
The harness — repryntt engine
An in-house autonomous framework that turns any frontier model into a worker: tools to act with, judgment gates before anything outward, persistent memory, and an append-only receipt for every action. Model-agnostic by design — swap the brain, keep the worker.
receipt · Runs our own company daily and every customer deployment; the same harness drives our Jetson robot lab.
02
shipped · revenue
The workforce — AI employees
The flagship product: managed AI employees at $199/mo doing named jobs — Front Desk, Outreach, Content — with 21 bench specialists carrying 158 trained deliverables behind them, and new deliverables shaped to order. One employee owns a whole function, the way real hiring works — and every employee is demonstrable live before a dollar is paid.
receipt · Live demos on every employee page; a public ops feed; receipts a customer can audit.
03
shipped · certified runs
The Distillery — micro-model refinery
Work leaves traces; traces are training data. The Distillery compresses a role's bounded decisions into ~10M-parameter specialists — certified against held-out data, gated (fail closed), and published with the teacher-vs-student delta. A specialist ships at roughly 12MB and answers in milliseconds on an 8GB edge board.
receipt · Certified reflex specialist: 90.3% holdout, 12.3MB artifact, 36ms p50 on a Jetson Orin — 9.7× faster than its teacher, −5.2pp accuracy, published.
04
research direction
The substrate — edge & embodied
A 12MB dense specialist is the natural payload for the compute that's coming: edge boards today, and the analog / in-memory / thermodynamic substrates being researched industry-wide. The endgame we build toward: certified specialist minds as the plug-in brains of working machines.
receipt · Today's receipt: the harness and distilled models already run on an 8GB Jetson in our own robot lab. Everything past that is labeled research, on purpose.
the problems we're working
Hard problems, honest approaches.
problem 01
One big model can't run a machine.
Frontier models are brilliant and slow; real work — a phone mid-ring, a sensor mid-spike — can't wait for a data center. Our approach: split cognition from reflex. Tiny certified specialists make the fast, bounded decisions locally; the frontier model is reserved for judgment. Cost and latency fall; competence doesn't.
problem 02
Small models lie about their competence.
Compress a model and it will happily keep answering — wrongly. Our approach: certification gates. Held-out evaluation, behavioral probes, head-to-head against the teacher judged by an independent model, and a gate that fails closed. Every specialist ships with its deltas published, including the losses.
problem 03
Machines don't speak in tokens.
Sensor streams — depth, inertial, thermal — are continuous signals; forcing them through a language model destroys the latency that keeps a machine safe. Our approach: language models supervise, they don't steer. Continuous control belongs to dedicated low-latency policies, with the language tier deciding what, never how fast.
problem 04
Autonomy without accountability is a liability.
An AI that acts in the world needs more than good intentions. Our approach — the oldest doctrine in the company: an append-only receipt for every action, approval gates on anything outward, and a human who can take over at any moment. Trust is an artifact you can audit, not a promise.
We publish problems and results — certified numbers, including the regressions. The recipes stay in-house. That's the line, and it's deliberate.
The loop that makes it compound
Work leaves receipts → receipts become each role's craft → craft and traces distill into certified micro-models → cheap local reflexes buy more work. The employee a customer has in month three is sharper than the one they hired — and the company's exhaust is a growing library of tiny certified minds.