The lab
We find out if it works before you pay for it to be built.
Research is not a phase we bill around. It is the first thing we deliver: a written experiment, run on your real cases, with results you can read.
The experiment record
Written down before it runs.
Every experiment starts as a record like this one. The question, the success line and the kill rule are fixed before we look at a single result, so the outcome can’t be argued into a win.
- Every experiment states its question and success line before it runs.
- Every experiment has a kill rule, and we follow it.
- Failed experiments are recorded and reported, not hidden.
- Results are measured on real cases, not on hand-picked examples.
Voice booking agent: turn latency
- Question
- Can a voice agent book appointments without callers noticing a delay?
- Success
- p95 end-to-end turn latency under 800 ms; booking completed without a human
- Variables
- Speech models, LLM, streaming strategy, end-of-turn detection
- Test set
- Recorded real-world calls, including interruptions and background noise
- Kill rule
- If the target is missed on every configuration, we stop and report
Research areas
Four questions under everything we build.
-
Area 01
Agent reliability
What makes an agent a colleague you can trust with write access?
Planning, tool use, recovery from its own mistakes, and the guardrails that keep it inside the lines.
-
Area 02
Voice and conversation
Can a voice agent follow a real conversation, interruptions and all?
Latency budgets, end-of-turn detection, barge-in and hand-off to a person.
-
Area 03
Evaluation
How do you know it will behave in production?
Test sets built from real cases, regression checks in CI and metrics that match the business goal.
-
Area 04
Cost and speed
What does each task cost, and how fast can it run?
Model routing, caching and architecture choices measured per task, not per demo.
What leaves the lab
Three possible outcomes. All of them useful.
-
Drop
The success line held. We build the production version on the approach that won.
-
Iterate
Close, but not there. We change one variable, write a new record and run again.
-
Stop
The kill rule fired. You get the data and our reasons, before any build budget is spent.
A fast “no” is a result too.
Bring us the idea. We’ll bring the evidence.
Tell us what you want to build. We will tell you how we would test it, and what a first experiment would show.
Start a conversation