Things I've built
Each project started with a real problem — and turned into a system I actually use.
task-board.md Task Router Agent — State-machine orchestration on a Markdown file
Running daily on my own machine · Session-independent · Self-healing state machine
An LLM-powered task runtime where a single Markdown file is the control plane. Typed state machine + single-writer protocol + self-healing. The insight: structured state, not better prompts, is what turns agents from execution bots into reasoning partners.
AI Content Automation Pipeline
Running daily on my own content workflow · Methodology file self-updates via feedback loop
7 agents covering topic → script → hook → title → quality check → data review. Each agent has a typed I/O contract; the Data Review agent writes validated rules back into a shared methodology file that the Topic Filter reads on the next cycle — so the filter gets stricter over time without prompt changes.
Rooted (根) — Interaction-first design for Chinese character learning
Live prototype · Product-thesis validation · Interaction-first language-learning design
One high-fidelity interaction per version, each carrying a single design thesis. V1: slide through 3,000 years of 马 (horse) from Oracle Bone to modern strokes. V2: tap 妈 and watch it decompose into 女 + 马 (phonetic). One interaction, one thesis, no feature list.
Function Calling Demo — a minimal Anthropic tool-use loop
Public GitHub repo · ~350 lines of Python · 13 pytest tests (security boundary included)
A minimal tool-use agent written without any framework — direct Anthropic API. Ships 5 typed tools (each with a JSON Schema), a dispatcher, a path-traversal guard enforced at the handler level, and a MAX_TOOL_ITERATIONS cap with a comment explaining when it should fire. Built to understand the primitive, not the framework.
Energy Demand Forecasting & Supply Optimisation
Deloitte-mentored case study — university group project, Paris
End-to-end quantitative framework for South Africa’s electricity system: 6 forecasting models benchmarked (ARIMA → TFT) + Pyomo dispatch optimisation under carbon / capacity / contractual constraints.
Robotics Exposure Index for European Labour Markets
Big Data for Development — Group research, Panthéon-Sorbonne
Exposure index for European occupations to robotics adoption, via LSA on patent + job-description corpora. Middle-wage routine-task jobs face the highest automation risk (consistent with labour market polarisation theory).