Editorial shelf
A good first wander through Peter's current ambition, systems depth, and delivery.
An employer-IP-bounded case study on Aurum at SThree: enterprise agentic AI engineering, reviewable state, fairness diagnostics, red-team testing, production control, and the self-improving-evaluator research that moved scoring quality from a 0.627 baseline to 0.851 without learning the benchmark.
How Datamise built the equal-pay work behind GMB Union's campaigns — an owned, relational claims base and the 38-finding audit that hardened it, plus SettleMise, a valuation engine where every number is traceable and a reviewer can take it apart.
Editorial shelf
Aurum, scoring systems that learn to score better over time, and the cost trade-offs of running AI in production at SThree.
An employer-IP-bounded case study on Aurum at SThree: enterprise agentic AI engineering, reviewable state, fairness diagnostics, red-team testing, production control, and the self-improving-evaluator research that moved scoring quality from a 0.627 baseline to 0.851 without learning the benchmark.
A 30-experiment programme that replaced a per-skill LLM validator with a zero-LLM retrieval-and-filter cascade and a 7 kB classifier head — 180× cheaper, ~100× faster, and within ~2pp of the model it replaced.
Editorial shelf
Datamise client work and reviewable equal-pay valuation software.
Editorial shelf
Oxford medical AI, Astroscale satellite vision, and turning simulated images into realistic ones.
A technical account of my Oxford DPhil: 3 million unlabelled cell nuclei from mouse placental histology, an 89% supervised classifier, near-perfect clustering of 8 cell types, and a self-supervised experiment that fell short of its goal.
How I approached spacecraft pose estimation at Astroscale: measuring the sim-to-real gap before modelling it, a preprocessing result that flipped sign with placement, the SPS-B architecture proposal, and AstroGAN — an image-translation model that closed the appearance gap while leaving the harder half open.
Editorial shelf
Benchmarking, evaluator research, analytics, public-sector strategy, and creative AI.
A controlled benchmark of 13 deep-research agent architectures on one shared tool layer — what happens to orchestration claims when you fix the model, the tools, and the judging protocol, and read the statistics honestly.
A technical log of an AI-assisted family film: colour-managed restoration, image-to-video animation, identity preservation, and the local pipeline that got built and then bypassed.
