Research
Compiler research. Engineering results.
How we search for optimizations, prove that behavior is identical, measure performance on real workloads, and run dedicated inference for teams.
-
01
EnterpriseDedicated inference your security team can review
Where requests go, what operational data is kept and for how long, how keys and team roles are scoped, and how failures are reported, with the limits of each answer stated.
Read the article → -
02
ProductPay for capacity, not tokens
Why dedicated endpoints are priced by time instead of tokens: your own hardware, no token quota, a price for the time it runs and spending limits set before the first request.
Read the article → -
03
ProductMore tokens from the same hardware
Inference spend tracks GPU capacity. How autonomous optimization across the stack gets more tokens from the same hardware, with the original model as the reference for quality.
Read the article → -
04
ThesisThe performance is already in the machine
The industry is spending hundreds of billions on new compute while leaving performance inside the hardware it already owns. What the published evidence says, what one small change did on an ESP32-S3, and why agents need a loop before they can do performance engineering.
Read the article → -
05
System designEvo: automating performance engineering
Inside the system we are building: two kinds of optimization search, independent checks and measurements, and an early liblc3 decoder experiment, including the upstream feedback that a win on one machine is not a win on every machine.
Read the article →