I’m Max Grody, a data engineer and applied machine learning researcher in Rockville, Maryland.
By day I build production data pipelines at Mathematica, mostly over Medicaid claims in Spark and Databricks. On my own time I build things over corpora I care about: a retrieval system that is designed to surprise as well as answer, a set of tools over the full transcript archive of a podcast I have listened to for years, and open-weight language models with layer-level surgery.
The common thread is that I would rather measure a thing than assert it. Each project below ships with its own evaluation, and when the evaluation said a version was worse, that version did not ship. Most of this work was done with Claude as a collaborator, and I try to be upfront about that wherever the work appears.
This site is where the projects live and where I write about how they were built. If any of it is useful to you, or you are working on something adjacent, I would be glad to hear from you.
GitHub · Hugging Face · Resume (PDF) · Email · RSS
439 episodes · 7.4M words · live
The full transcript archive of Sean Carroll’s Mindscape, read as one body of text. Topic models at three resolutions, a serendipity recommender, guest and time analyses, and a search that understands the question and returns the moment with a timestamp. Built from a 16-stage pipeline that rebuilds only what changed.
2,581 listener questions · 63 AMA episodes · live
Type a question and see whether the host has already answered it, and where. Semantic matching over every Q&A pair in the show’s AMA episodes, about 150 ms per query, no language model in the loop.
Retrieval with a certified surplus · MCP server · GitHub
A local retrieval system for Claude agents with a guaranteed-relevance floor (dense pool, lexical recall, cross-encoder rerank) and, above it, one statistically certified non-obvious passage per query, chosen by a sparse-autoencoder concept space and PMI certificates against the corpus citation graph. Serves the Stanford Encyclopedia of Philosophy and the Mindscape archive.
7 open-weight MLX releases · ~3k downloads/month
Qwen3.8-32B-Jumbo restores knowledge lost in continued training by transplanting the most-drifted layer blocks from a sibling checkpoint, chosen by a per-tensor alignment census; Mini-Me is a sheared 23B. The LoRA toolkit behind them measures catastrophic forgetting on every run, and the fine-tuned variants that regressed were not released.
A short note on the machinery behind preferencespace.com, which doubles as the example post for the format.
A build log for the tools over 439 episodes of Sean Carroll's Mindscape, from a respectful scraper to topic models, a serendipity recommender, and a search that understands the question.
An essay from the first months of the pandemic on the ethics of digital contact tracing, with a note from five years on.