ARTICLES
Notes on AI and software.
Essays about practical AI systems: how they are designed, where they break and what makes them genuinely useful to the people who rely on them.
EDITORIAL NOTE
Technical ideas, written for people.
This is a growing editorial archive. I will add more project notes, technical breakdowns and essays here as my work evolves. The writing is grounded in public project work and a simple conviction: strong technical systems should be understandable, reviewable and designed around the people who use them.
LATEST ARTICLES
9 articles · bilingualArchitecture · Reliability · 5 min read
Retrying is easy. Avoiding duplicate actions is harder.
A timeout tells you that the reply did not arrive. It does not tell you whether the action happened. That distinction should shape every workflow that writes to an external service.
Retrieval · Evaluation · 6 min read
RAG: evaluate the evidence before the answer
A fluent answer with three citations can still use the wrong document. To improve a retrieval system, first find out whether the failure happened before or after the model saw the evidence.
Agentic AI · 9 min read
Building agentic AI beyond the demo
The demo of my newsletter agent worked the first evening. What broke over the following week had nothing to do with the model, and everything to do with what travelled between the steps.
Edge AI · 8 min read
Designing AI that works without a network
Two Android projects taught me that on-device AI is mostly a budgeting problem. The model is chosen last, after memory, load time and the cost of being wrong have been written down.
Human-AI interaction · 8 min read
Making AI uncertainty visible
A confidence of 0.61 on the word “two” in “give him two tablets” and a 0.61 on an “uh” should produce opposite interface behaviour. Most products show them the same way, or not at all.
AI systems · 9 min read
The AI harness: the system around the model
During the Mistral hackathon we ran the same model through two versions of the Vibe CLI. One version completed tasks reliably; the other wandered. The difference was entirely in what surrounded the model.
Project note · OfflineLingo · 8 min read
OfflineLingo: designing for the last mile of communication
Built over a weekend at the European Defense Tech Hackathon in Berlin: an Android app that turns a spoken phrase into a checkable translation with the network permission removed from the manifest.
Robotics · Project note · 9 min read
Robotics is a chain of decisions, not a single model
At the AMD Open Robotics Hackathon our team, CRC, built a voice-controlled arm that picks up a camera, frames a person or object and keeps them in the shot as they move. The detector was the easy part.
Neurotechnology · Time series · 12 min read
Brain–computer interfaces are time-series systems before they are AI demos
An EEG electrode sees a few microvolts of summed cortical activity through skull and skin. Turning that into a reliable command is a signal-processing and evaluation problem long before it is a model problem.
ARTICLES