Decisions, written down.
Architecture Decision Records from the work we ship. Context, decision, consequences. Engineering writing — not marketing.
- ADR-00520 May 2026 · 3 min
Hallucination and data leakage: how we treat them as engineering problems, not vibes
Two failure modes — making things up, and leaking what shouldn't leave — show up in every production LLM system. We treat both as defects with measurable rates, not character flaws of "the AI".
llmsecuritygdprdiscipline - ADR-00129 April 2026 · 2 min
Pseudonymising CV + JD before any LLM call
A token-swap pipeline replaces every PII field with a deterministic stand-in before payloads leave our trust boundary. The model never sees a real name. Cost: ~80ms p95.
llmprivacygdpr - ADR-00215 April 2026 · 2 min
pgvector over a dedicated vector DB
Adding a second data system for one use case was operational overhead we'd regret. pgvector inside the Postgres we already had is good enough for the next three orders of magnitude.
datainfrastructuretradeoffs - ADR-0033 April 2026 · 1 min
Why we shipped Walkie-Talkie before Live Conversational
Turn-based audio ships in a week with no realtime infrastructure. Live Conversational mode is a month of WebRTC for a feature most users would pay for separately. Sequenced deliberately.
productscopesequencing - ADR-00418 March 2026 · 1 min
No social, no leaderboards, no streaks
The competitors are addictive-by-design. Celer deliberately isn't. The cost is lower DAU. The outcome is a defensible position.
productprinciples