Working notes on infrastructure, self-hosted models, and building software for people who are unwell. Written by the people doing the work, mostly while it's still fresh.
Before we chose where inference would run, we wrote down the threat model: who can see a message between the moment someone types it and the moment they get a reply. The answer ruled out most of the convenient options. Here's the full reasoning, the hardware it committed us to, and what we gave up to keep the promise.
Read the postStreaks, badges, and red notification dots are motivational tools built for people who already feel fine. What we replaced them with, and how the app behaves when a user disappears for three weeks.
2026 · 06 · 09 A backup is a restore you haven't tested yetOur restore drill, start to finish: what we broke on purpose, how long it took to come back, and the three assumptions that turned out to be wrong.
2026 · 05 · 22 Running several models on one box without them fightingMemory budgets, scheduling, and the quantisation trade-offs that let a modest server serve more than one workload at a usable latency.
2026 · 05 · 03 Hosting an open-source social platform, and what moderation really costsPeer support is genuinely useful in behavioral health. It's also a moderation problem you inherit on day one. What we planned for, and what we didn't.
2026 · 04 · 18 The evaluation set we wrote before we wrote the featureHow we test whether a model response is actually helpful to an anxious user, when "sounds supportive" is the easiest thing in the world to fake.
2026 · 03 · 30 Ubuntu LTS, boring on purposeWhy a small team should choose the least interesting base OS available, and the upgrade calendar we keep so that choice stays cheap.