Blog
Notes on AI, ML, and what runs them.
Fine-tuning, retrieval, agents, inference, evaluation — and the infrastructure underneath all of it. Written from things I've actually shipped, including what didn't work.
Nothing published yet.
The first posts are being written. Until then, the code and the models are the best look at what I'm working on.
In the queue
What I'm planning to write about.
- Fine-tuning open models LoRA and full fine-tunes, data preparation, and when tuning actually beats prompting.
- RAG that survives real data Chunking, embeddings and reranking once the corpus stops being a tidy demo set.
- Agents past the demo stage Tool calling, failure handling, and keeping cost and latency inside something usable.
- Serving and inference Throughput, batching and quantization — getting more out of the GPUs you already have.
- Knowing whether it got better Evaluation that catches regressions instead of vibes-checking a handful of prompts.
- The cluster underneath Power, cooling, InfiniBand and solar — what running 32 GPUs actually involves.
Something specific?
Tell me what you'd want written up.
If there's a part of this you're stuck on yourself, that's a good reason for it to be the first post.