Selected writing

Writing

01 / WRITING

Sep 2026 12 min read

Don't call .all() on a training export

Why a generator that opens with .all() isn't lazy at all, and the three changes — streaming with yield_per, offloading CSV formatting to Postgres COPY, and batching the rows Python still has to touch — that let a recommender's training export survive tenants with millions of users instead of getting OOM-killed.

Jul 2026 10 min read

A recommendation you can't explain won't get used

Why every career recommendation we produced shipped with a plain-language "why", how we extracted explanation signals from the model's own internals, and the honest ways those explanations can still mislead.

Jul 2026 9 min read

One tenant, one shard — hiding Postgres sharding from the whole app

How I spread a platform's customers across several Postgres databases instead of one, and hid the routing behind a session helper so the rest of the code never has to know a shard exists.

Jul 2026 8 min read

I didn't need a lock server — Postgres already was one

How I stopped two copies of the same service from ever running one tenant's pipeline at the same time using a single Postgres function call, instead of standing up Redis or ZooKeeper — and why a crashed copy never leaves a stale lock behind.

Jul 2026 10 min read

Uptime says green; the pipeline is broken

Unit tests prove the parts work and an uptime ping proves the server answers, but neither proves the whole pipeline still turns raw data into a real recommendation — so I built a watchdog that quietly runs a fake customer through it end to end, on a schedule.