Notes from production AI

How the Speculo and Smart Advocate systems were built, and how they held up once live.

• 117K+ emails and SMS, plus about 10K AI voice calls, in 6 hours on GCP Pub/Sub

  Rate limits, at-least-once delivery, and the idempotency keys that stopped duplicate sends.

  Apr 2026 · 14 min read

• The agent backend behind Speculo's outreach

  Four LangGraph agents, one delivery service, and why state management was the hard part.

  Apr 2026 · 7 min read

• KAG vs RAG: why we switched to knowledge-augmented generation for legal AI

  Where standard RAG struggled on legal case files, and how we measured the 40% gain.

  Apr 2026 · 8 min read

←Back to work