/now

This week

Cross-video pattern detection in Retention Lab. Building the thing YouTube's native analytics won't: grouping retention curves across an entire channel instead of staring at them one video at a time.

Also in flight:

• Tightening the Mnemix warm path — caller memory is under 300ms on the edge; chasing the cold-path number next.
• heycli: reducing idle cost on the orchestrator after I found it burning ~$14/day doing nothing.

What's next

• Eval harness for Retention Lab's clustering — a real metric on “did we surface the right pattern,” not vibes.
• Writing up the BEAD model (bi-temporal, evidence-anchored decisions) as a public post.

Recently shipped

• The flagship postmortem: The night the doctrine failed. Four patterns committed to MOLL.
• Sub-300ms caller memory on the Mnemix warm path (Cloudflare edge → Upstash Redis).
• heycli pushing prompts to Claude Code, Codex, and Aider from one source-of-truth.