back to home

writes

Breaking the NN Ceiling — LightGBM, FT-Transformer, and Why Ensembles Work

#machine-learning#lightgbm#transformer#ensemble#onnx

The neural net was stuck at 264s. Diagnostics said rare pairs were the problem. The fix wasn't a better NN — it was two completely different models and an ensemble that cut MAE to 253s.

From 351s to 253s — Feature Engineering, Neural Nets, and Why Statistics Beat XGBoost

#machine-learning#feature-engineering#pytorch#deep-learning

Zone-pair median — a dictionary lookup — beats XGBoost with zero ML. Here's how I built 26 features, iterated through 4 neural net versions, and hit a ceiling that no architecture change could break.

From 0.300 to 0.537 — 15 Experiments Training an LLM Stock Trader with SFT and GRPO

#reinforcement-learning#llm#sft#grpo#deepseek#fine-tuning

15 models. 4 disasters. A complete experiment log of training a 7B LLM to trade stocks using supervised fine-tuning and Group Relative Policy Optimization — what worked, what broke, and why.

Building an RL Environment That Actually Works — Rewards, World Models, and Why Environments Are Harder Than Training

#reinforcement-learning#llm#pytorch#openenv#world-model

Every shortcut I left in the environment, the agent found and exploited. Here's how I designed observations, rewards, grading, and a neural world model for training LLM trading agents on real Indian equity data.