Journal
Engineering writing on the things we actually build: vision-language-action models and robot autonomy, multi-agent orchestration, retrieval that survives a real corpus, and the rendering work underneath experiences like this site.
Embodied AI
A 3B-parameter policy cannot run at 50 Hz, and yet dexterous manipulation needs 50 Hz. Action chunking is how that contradiction gets resolved — and temporal ensembling is how you pay for it without the seams showing.
Embodied AI
Denoising an action sequence is an unreasonably good way to model multimodal behaviour. It is also a sampling loop inside a control loop, which is where the engineering starts.
Embodied AI
Robot data is not scraped, it is manufactured — one operator, one episode at a time. The teams that win at imitation learning are the ones who treat data collection as a production system rather than a chore.
Embodied AI
Simulation gives you unlimited data and a reset button. It also gives you a physics engine that is wrong in specific, characterizable ways. Knowing which ways is most of the skill.
Embodied AI
Photons hit the sensor; some milliseconds later a joint moves. Everything in between is a budget you either write down and defend, or discover the hard way when the arm oscillates.
Embodied AI
A learned policy cannot be verified. A safety layer around it can. The engineering problem is making the layer restrictive enough to be meaningful and permissive enough that the policy can still do the task.
Agent Systems
Most multi-agent architectures are one of five shapes. Picking the wrong one costs you tokens, latency, and debuggability — and a surprising number of problems are better served by one agent with better tools.
Agent Systems
A long-running agent's context window is a resource with an allocation policy, an eviction strategy, and a failure mode. Treating it like a scratch buffer is why long runs degrade.
Agent Systems
Most agent failures attributed to reasoning are really interface failures. The tool surface is a user interface, the user is a language model, and it has specific and predictable weaknesses.
Agent Systems
A benchmark score tells you how a model does on someone else's problem. Building an evaluation harness for your own is unglamorous, statistically fiddly, and the highest-leverage thing most teams are not doing.
Agent Systems
Vector search on a demo corpus is a weekend. Retrieval over 400,000 real documents — tables, near-duplicates, acronyms, twelve years of revisions — is a systems problem that embeddings alone do not solve.
Machine Learning
If a policy could imagine the consequences of an action, it could plan instead of react. Learned dynamics models make that possible — and introduce a new way to be confidently wrong.
Embodied AI
Picking things up looks solved in demo videos. The gap between a demo and a system that works on an unseen object, in bad light, next to five other objects, is mostly a perception problem.
Real-time Graphics
Every object floating on this site is a mathematical function, not a mesh. Here is how sphere tracing works, why it suits the browser, and what it costs when you put a dozen of them on one page.
Let's build
If any of this is close to what your team is building, we'd like to hear about it — including the parts that aren't working yet.