Pinned Loading

1. rlhf-from-scratch-on-distilgpt2 rlhf-from-scratch-on-distilgpt2 Public

   RLHF pipeline from first principles: SFT, LoRA, reward modeling, PPO, and DPO/IPO/KTO/ORPO/SimPO, unit tested and run end to end on DistilGPT2

   Python

2. llm-agent-calibration-eval-harness llm-agent-calibration-eval-harness Public

   LLM eval / RL-environment design harness: calibrating agent failure modes across 5 forecasting tasks

3. scaling-llms-distributed-systems scaling-llms-distributed-systems Public

   Study notes on scaling LLM training, distributed systems, and inference, worked through the JAX scaling book end to end.

4. German-Product-Title-NER German-Product-Title-NER Public

   Production-grade multilingual NER system for structured attribute extraction from German e-commerce titles using XLM-R + CRF with modular training, HPO, and post-processing pipelines.

   Python

5. Kenda_Tires_Sales_Forecasting Kenda_Tires_Sales_Forecasting Public

   End to end SKU level sales forecasting system built for a tire distributor. LightGBM global model, Croston SBA for intermittent demand, MinT hierarchical reconciliation, and A to D confidence gradi…

   Jupyter Notebook

6. Shell-Machine-Learning-Fuel-Blend-Prediction Shell-Machine-Learning-Fuel-Blend-Prediction Public

   Developed an end-to-end ML pipeline for fuel blend property prediction using advanced preprocessing, blend-weighted features, nonlinear transformations, PCA, and clustering. Built a LightGBM + Ridg…

   Jupyter Notebook