Briefs on the systems built in the lab — agentic AI, governance, robotics, forecasting — alongside essays on how AI is reshaping learning, organisations, and software.
Each brief is a short read — the idea, the interesting engineering, and what the pattern is good for beyond the project itself.
ZenInvest let a committee of Claude, GPT, and Gemini debate every trade on a paper-trading account for six months. The honest answer was no. This brief covers what the evidence showed, why it was published, and how the redesign, ZEN-2, is built to measure LLM value rather than assume it.
ZenArena puts one of agentic AI's favourite claims — that memory makes agents better — under a falsifiable test, using chess and a Stockfish truth signal to measure whether governed memory beats remembering everything.
ZenRate wraps a deterministic actuarial core in a layer of collaborating AI agents — where exactly one agent holds a veto, every decision is recorded bi-temporally, and the compliance paperwork falls out of the architecture.
ZenForecast is an open-source, governance-first Python framework that treats forecasting as a continuous predict → decide → act → learn loop — with interchangeable models, leakage-free backtesting, and an audit trail behind a single API.
ZenLab is an offline-first research monorepo spanning 21 AI/ML topics across 10 tracks — a place where agentic techniques must earn their way into production designs through experiments instead of hype.
Stitch is a simulation-first digital twin of an autonomous, wildlife-safe, amphibious litter-collection robot — a six-package ROS 2 / Gazebo system that must pass its full mission in simulated physics before a single part is bought.
ZenGrowth scores senior AI and data-science roles with an explainable fit score and drafts applications grounded in verified evidence — locally, auditably, with deterministic gates against selected fabrication and a human reviewing every draft.
A solver-free connectivity test, grounded in the graphic matroid, can screen power-grid topology actions far more cheaply than simulating each one. I pre-registered a study to see whether that screen also helps a reinforcement-learning agent learn faster. It did not, and the reasons are worth more than a positive result would have been.
The focus is on work that helps people think more clearly about how AI systems are built and used in practice. No invented metrics, no hype — honest caveats stay in.