Independent research and open-source experiments in deep learning, LLMs, agents, and software systems.
A training framework that improves agent harnesses while keeping the LLM frozen, updating prompts, tools, and context management in a training loop. End-to-end determinism enables credit assignment to harness changes. Trained on a 35B open model, the harness transferred to six other held-out models on SWE-bench and Terminal-Bench 2.0 and outperformed the official Terminus 2 harness.
Python, Agent Systems, Harness Engineering, LLM Evaluation
A deep learning library built from scratch, from autograd to pretraining GPT-2 124M on 27B tokens and fine-tuning with SFT and GRPO. A PyTorch-like API with tensor operations, optimizers, and backward-pass math derived in code and validated against PyTorch. Trains CNNs, RNNs, and transformers on NumPy, MLX, or CuPy backends.
Python, NumPy
A real-time game-playing agent built with computer vision and reinforcement learning.
Python, TensorFlow, Reinforcement Learning, Deep Q-Learning, Computer Vision, Convolutional Neural Network
Abstractive Text Summarization using two approaches: (1) GRU+Attention (2) Transformers
Python, PyTorch, NumPy, Natural Language Processing
Research and experiments on group sparsity techniques to improve generalization and efficiency in deep learning models.
Python, Deep Learning, ML Theory
A collection of reinforcement-learning experiments across classic control environments, exploring policy-gradient and actor-critic methods.
Python, Reinforcement Learning, Policy Gradient, Actor-Critic