Notes from the practice.

On May 30th I published v0.1 of a Databricks integration pack for Swamp. By the end of the same evening it was at v0.13. Fifteen models, 100% A on every release, every model end-to-end smoke-tested. What shipped, what I learned, and why it matters if you have Databricks alongside anything else.

14 min read

Industry never required philosophical certainty before deploying complex systems. The AI deployment debate is no different. A case for treating intelligence as uncertainty compression.

4 min read

A postmortem of fine-tuning an 8B model for U.S. tariff classification on a single rented H100. Sixteen hours, $43, and eight lessons earned the hard way.

14 min read

Your Databricks cluster says 28 GB. Your Python code gets maybe 12. The JVM/Python split that explains why, and how to budget the rest.

7 min read

Feature engineering didn't die. It became lossy compression for language models, turning thousands of raw rows into ten metrics that carry the same story.

9 min read

Why public data deserves public infrastructure, and how patents on basic ML techniques applied to taxpayer-funded datasets are gatekeeping disguised as innovation.

17 min read

Four kinds of blind spots in AI deep research outputs, and a working mental model for using these tools as starting points rather than finished products.

7 min read

How Shazam identifies one song out of 11 million in 5 seconds, by refusing to solve the actual problem.

4 min read

Long before vector databases, Merriam-Webster lexicographers built a 315,000-card index of words spelled backwards. The same insight powers modern AI: different organizational schemes reveal different patterns in the same data.

8 min read

How one 2017 paper on machine translation accidentally became the architecture beneath ChatGPT, Claude, Copilot, DALL-E, and almost every other AI system you use.

9 min read

Max Planck discovered nature comes in discrete packets in 1900. 125 years later, the same idea lets vector databases, LLMs, and on-device AI work at scale.

6 min read

Runtimes make knowledge work executable. But isolated execution doesn't compound. The semantic mesh is the layer that connects structured outputs into accumulated organizational intelligence.

13 min read

Agentic AI thrives in software engineering because code already runs inside a repeatable execution environment. Most knowledge work doesn't. The ceiling isn't model intelligence, it's the absence of runtimes that can execute, validate, and provide feedback on structured work.

14 min read

When an agent makes a decision, you should be able to see exactly what it saw, and whether reality has changed since. Introducing QPS, a companion spec to TMS.

6 min read

Dashboards were built for eyes. LLMs need something else: a structured, multi-resolution data format optimized for machine cognition. Enter Tabular Manifolds.

6 min read