./vla-book-picking
2026Fine-tuning ACT and SmolVLA to pick colored books off a shelf in MuJoCo on one 5070 Ti. Peaked at 60%, then found out the model had memorized shelf slots. Mostly a negative result, written up in full.
things i've built, roughly grouped. most of the source is on github.
ls ~/projects/robotics-ml
Fine-tuning ACT and SmolVLA to pick colored books off a shelf in MuJoCo on one 5070 Ti. Peaked at 60%, then found out the model had memorized shelf slots. Mostly a negative result, written up in full.
Pipelines that turn egocentric human video into VLA training data, plus a harness to measure whether a processing choice actually helps the policy. Splits GPU work from cheap CPU ablations.
Kubernetes-native orchestration for robot inference. Robot, Model and Fleet CRDs, an operator loop, and an edge agent that syncs and loads models on the robot.
Fantasy Premier League lineups solved as an MILP across multiple prediction horizons. Bolting on an auto-research loop so it can swap in better predictors over time.
Undergrad thesis. Reconstructs a 3D model of a prompted object from a single image by chaining GroundingDINO, SAM, ZoeDepth and MCC. Shipped before Shap-E came out.
ls ~/projects/systems
A fast C++17 Sudoku solver. Constraint propagation plus backtracking, with 46 tests because at some point it became a craft project.
A calculator in Rust with a CLI and a GUI. Mostly an excuse to learn egui.
A breadboard circuit simulator with a Go backend and a web frontend.
ls ~/projects/older
A vanilla PHP cryptic-hunt platform for a school fest. 1000+ users, zero downtime. That it worked is still slightly miraculous.
A tiny decentralised forum where people could roast me and earn some (testnet) Ether.