Projects
Knowledge system
A personal knowledge system modelled on the brain. My notes are organised into seven “lobes”, with an LLM maintaining the knowledge layer and Notion as the display layer. It’s the foundation the other projects are built on, and the structure Emere’s experts are trained around.
Language model
A small language model I’m building from scratch: my own transformer architecture and tokeniser, no pre-trained weights. I’ve scaled it from 17M to 305.6M parameters across seven recorded training runs, plus an 8-expert mixture-of-experts version (linked to the Mind OS structure), documenting every run, plateau and failure along the way.
Agents
An organisation run by 19 AI agents, each with a defined role, reporting lines and rules to follow: hiring exams, an AI usage policy and a written reasoning record for every output. It runs a daily pipeline on locally hosted models, entirely in dry-run with no real money involved. It’s my sandbox for seeing how agents behave when given real roles and real constraints. The real test comes next: factoring in human attitudes to risk, to turn a generic finance agent team into one that takes and manages real risk.
Verification
An open-source framework for checking whether AI agents can be trusted. I put agents through hiring-style exams and found a failure that repeats across various models: when a prompt implies the data already exists, models invent it rather than admit they don’t have it. Thirteen exams on three local models, every conversation published (including the round that failed outright), and a guard that caught a real fabrication in 10 of 10 adversarial tests.
Research
A research programme I designed that an AI agent runs. On my command, it draws a topic at random from four areas (AI, SaaS, IT Procurement, & Technology) and works one genuinely disputed question into a dissertation-style paper, held to rules I set on sourcing, structure and fact-checking. Research and writing by an AI agent; created and directed by me.