Writing from the workbench.
What we are shipping, what we are learning, what we would build differently. Written from the desk, not the deck. New posts weekly.
Why we ship Voice AI in 4-6 weeks, not 4-6 months
Turn-taking sub-350ms is table stakes. Warm human hand-off is table stakes. What makes the difference between a demo and a working production agent is the eval loop — and how tight the scope was on day one.
Where GCP earns its price and where it doesn't
Vertex AI, Cloud Run, BigQuery, Firestore, Pub/Sub. Where each earns its cost line and where it is a tax you pay for convenience.
Multi-agent handovers · how CALEL Legion routes work between LLMs
Handover ladders, state monitors, and how to stop drift between agents built on different LLMs.
Aquaman · page-fill ratio as the real diagram moat
Why most AI-generated diagrams waste 60-70% of the page. What we do differently. Preview of the v2 dashboard engine.
SOS.VU · what personal safety software owes you
Why we built SOS.VU around the moments that actually matter to a person on the wrong end of an incident, not the moments that look good in a demo.
Wearable safety · the case for days of standby
Why we hold every wearable to the same bar — run all day without a charge, private by default, and only awake when it matters.
Working on something similar?
If a post here maps to a real problem you have, tell us. We reply within two working days.
Talk to us