a 351m-parameter gpt trained from scratch on 5b tokens, then reasoning-tuned with sft + grpo (rlvr) on math — full pretrain → sft → rlvr pipeline, end to end on a single amd mi300x, for under $100 total.

python · pytorch · rocm · grpo · hugging face

local-first agentos for minimal hardware — routes everyday queries to a free local gemma 4 model and escalates coding/reasoning work to cloud tiers, with multi-agent dispatch, persistent memory, cron jobs, and crash-recoverable event delivery.

python · fastapi · litellm · docker · gemma · fireworks ai

gpt-2 small (124m) trained from scratch on openwebtext — a file-per-concern reimplementation with flash attention, torch.compile, and bf16. 5k iterations on a single l4 gpu, val perplexity 27.3 against ~22.4 for the fully-trained original.

python · pytorch · flash attention · torch.compile · hugging face

fine-tuned qwen3.5-0.8b with lora on the llm-lat/harmful-dataset to strip its refusal behaviour — a red-teaming artifact for studying how fragile alignment is under targeted fine-tuning. held-out test loss 1.345, perplexity 3.84, on ~2 minutes of t4 compute.

python · lora / qlora · unsloth · hugging face · pytorch

mcp server exposing google stitch's ui generation api as structured tools — create projects, generate screens from natural language prompts, manage design systems, and produce design variants. lets claude and other ai clients drive interface creation conversationally.

python · mcp · google stitch api · fastmcp

mcp server exposing 40+ gitlab operations as structured tools for ai clients like claude and cursor — repo management, branch protection, merge requests, issue tracking, ci/cd log retrieval, and cross-project search. includes a standalone gemini-powered autonomous agent.

python · mcp · gitlab api · pydantic · docker