Three Rebuilds of This Website
How my personal site went from a Jekyll academic template, to Next.js and Tailwind on Vercel, to Astro on Cloudflare — and what each rewrite was really about.
Web Astro Meta
Writings
Autoresearch-Charged Action Search
How we used autonomous experiment loops to take action search from hand-crafted heuristics to a 109M model that beats embedding models 70x its size.
AI Search Engineering
Supercharging Research with Agents
How I built an end-to-end research system at a talk, from context files to a second brain with knowledge graphs.
AI LLMs Engineering
Computers and Computation
Teaching a 2B parameter model to delegate arithmetic to an external calculator via inline computation tokens, with three eval backends on Apple Silicon.
AI LLMs Engineering
Coding on my Phone
My journey building a mobile coding setup - from SSH terminals to custom apps to autonomous agents running overnight on a VPS.
AI LLMs Engineering
Building a Search Tool: Accelerating Development with AI
How I built semantic search for 10k+ actions at StackOne, and reflections on the AI-accelerated development lifecycle.
AI Engineering Production ML
Coding in the Future
My thoughts on where the evolution of coding in the world of LLMs.
AI LLMs
Building an Agentic Researcher at StackOne
How I built an AI agent that finds agentic use-cases and gets better with feedback.
AI Engineering Production ML
How I Serve My LLM So That My LLM Serves Me
Here is how I optimized the serving of my fine-tuned LLM to efficiently evaluate it on the Tau benchmark.
AI Engineering Production ML LLMs Inference
Principles of Speed: More GPU Memory Utilisation ≠ Better Performance
How to make your GPU utilisation go up by taking your GPU Memory usage How to make your GPU train an LLM fast(er)
AI GPUs
Panic Coding on My Phone in a Lightning Storm
How I spent the last day of my Dubrovnik holiday fixing my broken FPL bot while dodging lightning bolts and BBQ duties.
AI MCP FPL
Building an FPL MCP Server
How I built a Fantasy Premier League MCP Server to automate my terrible FPL career.
AI MCP FPL
Building an Eval-Driven AI Feature at StackOne
How I built an AI agent that translates provider errors into clear resolution steps, starting with evals from day one.
AI Engineering Evals Production ML
Time-Series Training Partitions
The training partition is a crucial component of an evaluation. It is used to assess generalisation performance on unseen data and ensure that the model is not overfitting to the training data. Here is a (probably) new way of partitioning time-series data for training and testing.
Evals Time-Series ML
WRandAi: How Much Does Your Benchmark Depend on the Seed?
A small PyPI package for quantifying how much a machine learning leaderboard depends on the choice of random seed.
Evaluation GNN Open Source
Branching Text Generation in Obsidian
An Obsidian Canvas plugin that turns LLM text generation into an explorable graph of possibilities.
LLMs TypeScript Projects
A Crypto Trader on a Raspberry Pi
Building a bespoke reinforcement learning trading agent with deep belief networks and sentiment data, running on a Raspberry Pi.
Reinforcement Learning Trading Projects
Controlling Spotify From the Wall
A hand-gesture music controller built from printed album art, a webcam, object detection and a 3D CNN.
Computer Vision CNN Projects