13 commits this week10 repos trackedlatest result 2026-08-08kvwarden v0.1.6 on pypi
systems engineering · applied ai
coconutlabs
A quiet tenant keeps its latency under load. I build the schedulers that make that true and measure them on rented hardware. The parts that failed are published too.
quiet-tenant p99 TTFT
flooder 32 rps · 300 s
1× A100 · Llama-3.1-8B · vLLM 0.19.1 · n=311 post-warmup · FIFO bar off scale. full provenance
ratio to solo
1.14×
vs fifo tail
26×
harness
public
engineers
two
work
Projects
KVWarden
1.14× of solo · 26× better than FIFO
Tenant fairness on shared inference. A quiet tenant stays near solo latency while a flooder pushes the same engine.
Project pagemlxd
Tenant-fair inference on Apple Silicon
A scheduler and admission layer on top of mlx_lm.server. Tenant identity first, then fairness on unified memory. Still in research.
Project pageCoconut OS
Agents as kernel primitives
A Linux distribution built on the position that the kernel should know what an agent is, what it may touch, and what it did. Spec phase, with kernel prototypes.
Project pagekvwarden.org and coconutos.org are separate domains for now. Everything is one lab. The running demos live in the hall of demos.
evidence
Recent results
2026-08-08
A model in the room
Building a synthesizer plugin with an AI pair: three jobs the model has in a creative tool, and the fence that keeps it useful.
2026-08-08
What mixing taught me about evals
Judgment under unreliable perception is a solved craft. Music solved it with structure, and my synth now carries a 3,050,113-assertion regression suite to prove the transfer runs both ways.
2026-04-19
Tenant fairness on shared inference
KVWarden Gate 2: 53.9 ms quiet TTFT, 61.5 ms under flooder pressure, and 26× better tail behavior than FIFO.
Coconut Labs works on the shared layer of inference: scheduling, fairness, cache pressure, and the measurements that keep claims honest.
The lab is small by design. Fewer abstractions between the benchmark, the note, and the code.
1,585 ms of waiting for a prompt that costs 53.9 ms with nothing else on the box.
the estate
One lab, several surfaces
Five doors. Here is what is behind each one.
kvwarden.org
The scheduler that keeps a quiet tenant quiet. v0.1.6 on PyPI, harness public.
masterclass.coconutlabs.org
Five chapters on forward-deployed engineering, start to finish.
coconutos.org
Paused while the lab reworks what Coconut is. The page behind this door says so itself.
The book under Coconut OS
Nineteen pages from the transistor to the scheduler.
coconutlabs.org/live
The operations record: a closed ingestion run's gap ledger, and the nightly loop.
the lab
One engineer, close to the work.
Coconut Labs is the name I publish under. The work happens in the open at github.com/coconut-labs and shows up here when there is a result worth standing behind. Jay Patel has built alongside me on several of these and is credited on them.
Building something at this layer? Write us.