We build distributed training and inference systems, grounded in the fundamentals of machine learning and of the systems it runs on. When the frontier moves — and it moves often — we read the work, reproduce what we can, and form our own view before it touches production.
01 / Services
Autoregressive models sample every token — including the ones that must be exact. For analytical work we build systems where the model orchestrates rather than recites: it writes programs, an interpreter computes the facts, and generated prose carries them.
Every figure in the output traces back to a computation, not a sample. Reports can be verified programmatically, and failures are loud — a crashed program beats a confidently formatted wrong number.
A round trip to an inference API is not a function call. It is long-running, failure-prone, and inherently asynchronous — and every tool call is another one. We build platforms that treat this as the ground truth: event-driven pipelines, streaming by default, composed as reactive flows rather than blocking request chains.
Backpressure, retries, cancellation, and replay are designed in from the start — so agentic workloads degrade predictably under load and recover without human intervention.
Your models, on your hardware. We help teams stand up inference across their own machines — serving stacks, schedulers, and the networking beneath them — tuned to the models they actually run.
The case is usually cost, control, or data boundaries. We make those numbers explicit before you commit, and leave behind a platform your own team can operate.
02 / Approach
The field moves quickly, and the incentives favour noise. When a new training recipe displaces last year’s consensus, or an architectural shift rewrites the economics of serving, the useful question is never “is this the future?” It is: what does this actually change, and for whom? Answering that takes fundamentals — in the mathematics and in the systems — and the patience to reproduce results before holding opinions about them.
So we work in the open. Every engagement produces a system and a body of writing — field reports, reproducible experiments, and an honest accounting of what didn’t work. The understanding compounds for your team long after the engagement ends.
New results are claims. We reproduce and measure before we adopt — and before we dismiss.
We publish methods, benchmarks, and the failures — not just the wins.
Built for many machines from the first commit, not retrofitted later.
If we cannot measure it and explain it, we do not ship it.
03 / Writing & Experiments
All writingAutoregressive models sample every token, including the ones that must be exact. We split the output into two regimes: computed where variance is failure, generated where variance is fine, and let the interpreter act as the external critic that models provably cannot be for themselves.
