Pinned Loading

1. sft-crm-agent sft-crm-agent Public

   LoRA SFT of Qwen2.5-1.5B (trainer written from scratch): parts-broker inbox message -> CRM record + rule-checked draft reply. 97.0% vs 84.8% for gpt-oss-20b 6-shot on held-out test.

   Python

2. gpt-cpp-cuda gpt-cpp-cuda Public

   GPT training and inference engine written from scratch in C++/CUDA. Hand-written forward and backward kernels, cuBLAS matmuls, no PyTorch or JAX.

   Cuda

3. ml-jax-pytorch ml-jax-pytorch Public

   Canonical models, whittled down: baseline → optimized → benchmark. Rung 1: MNIST CNN + int8 quantization.

   Python

4. rl-lisp-python rl-lisp-python Public

   GRPO from scratch in JAX on one 8GB GPU: teaching a 0.5B code model to write Python as if it were Lisp. No trl, no peft, no vllm.

   Python