Machine Learning Projects
I
I built SupplyChainBench, an RL environment and benchmark designed to test whether LLMs can learn to make better decisions when the consequences of their actions only become visible several steps later. Using the Beer Distribution Game, models must forecast demand, account for orders already in the pipeline, and make decisions whose effects unfold over several weeks. On 16 held-out stochastic benchmark seeds, Qwen3.5-4B scored 38.42 with a GRPO + LoRA adapter versus 11.78 untrained—a 3.3× improvement. The adapter was trained on the earlier environment and evaluated on the stochastic version without retraining.
II
A robot works out what to do next by looking up the closest moments it has already seen. Retrieval over a demonstration memory cuts action-prediction error 6.5x against a direct baseline on 25 demonstrations.
III
Type a ticker, pick a DCF or an exit multiple, and get five-year projections with a fair value and the upside from today's price.
IV
A GPT written from scratch, tokenizer and attention and sampling included, then trained at three sizes to watch validation perplexity fall from 17.75 to 8.08.
Foundations
Earlier builds the current work rests on.
-
Connect 4
DQN and PPO implemented from the derivations and trained on Connect Four over a shared convolutional board encoder, so the learning rule is the only variable between them.
-
JEPA experiment
A controlled comparison of JEPA-style latent prediction against masked-pixel reconstruction on matched ViT backbones, now folded into RAID as lineage.
-
Project Aurora
A Bayesian scenario model for a multi-country manufacturing location decision, replacing point estimates with posteriors fit to FRED series.