01 /Research
Research on product engineering & AI systems.
Original research on full-stack craft, AI systems, automation, and the product decisions behind what I build.
Index · 08 entries
2026 — Present
- 01ExperimentOctober 20, 2026 (3d ago)
Why voice AI agents create duplicate bookings
A repeated confirmation should reuse one booking result. Here is how I test that behavior in my Voice Reliability Lab, and where production safeguards must go further.
5 min readvoice-ai · livekit · testing - 02Deep diveSeptember 12, 2026 (1mo ago)
Your voice agent heard the correction. The tool still got the old request
A natural conversation can hide a stale or unconfirmed tool call. Action integrity links the caller's latest intent to an explicit confirmation and one observable execution.
8 min readvoice-ai · agents · reliability - 03NoteSeptember 8, 2026 (1mo ago)
Before you buy an AI receptionist, audit 7 days of missed calls
Use one week of your own call history to decide whether after-hours or overflow call coverage is worth testing—without relying on generic revenue claims.
5 min readvoice-ai · hvac · operations - 04NoteSeptember 8, 2026 (1mo ago)
Your voice agent understood the call. Did it capture the customer?
Names, emails, phone numbers, and addresses are easy to hear incorrectly and expensive to save incorrectly. Here is how I would test customer-detail capture before a voice agent takes real calls.
6 min readvoice-ai · transcription · testing - 05NoteSeptember 7, 2026 (1mo ago)
When can an AI receptionist safely say 'you're booked'?
A calendar timeout can leave a booking in an uncertain state. Here is how I would test confirmations, retries, and dispatch handoffs before an AI receptionist takes real calls.
5 min readvoice-ai · hvac · testing - 06NoteSeptember 6, 2026 (1mo ago)
What to test before an HVAC voice agent takes real calls
A practical pilot checklist: service-area checks, uncertain scheduling, human handoff, failure handling, and the records dispatch actually needs.
4 min readvoice-ai · hvac · testing - 07ThesisMay 20, 2026 (5mo ago)
Voice context is a timing problem
Long voice calls slow down for a familiar reason: too much history, not enough context window. The usual advice is to summarize the old turns. That assumes you have time to run the summary. In voice, you don't. The fix isn't a better summary. It's deciding when the summary runs.
2 min readvoice-ai · agents · latency - 08ThesisApril 20, 2026 (6mo ago)
The Contact Graph: why agents need a social network, not a database
Agent memory today is a vector store — text chunks retrieved by keyword similarity. That breaks the moment you need to remember who said what, how reliable they were, or what they still owe you. Here's a different primitive: the contact graph. Memory modeled as a social network, with facts attached to relationships.
8 min readagents · memory · context-engineering