Coconut Labs is where @ShreyPatel4 publishes his own systems research. Independent, self funded, no investors and no roadmap owed to anyone. The work is mostly schedulers: who gets the next slot, how long they wait, and what it costs to be fair about it.
Nothing here ships a number that has not been measured. Where a claim is soft, it says so.
Tenant-fair LLM inference on a single GPU. No Kubernetes. A token bucket sits in front of vLLM or SGLang and admits per tenant, so one caller's flood stops starving everyone else on the same engine.
Measured on A100-SXM4 with Llama-3.1-8B and vLLM 0.19.1, two tenants, 300 s sustained, quiet-tenant TTFT p99:
| p99 | |
|---|---|
| Solo, no contention | 53.9 ms |
| FIFO under a flooder | 1,585 ms (29x starvation) |
| kvwarden token bucket | 61.5 ms (1.14x solo) |
Ten lines of YAML. No application code change. Raw artifacts ship in the repo.
pip install kvwarden
Fills the gap between Ollama, which assumes one user, and Dynamo or llm-d, which assume a datacenter.
A Linux distribution where agents are first-class kernel objects rather than processes that happen to call an API. Capability tokens, an audit path, and a scheduling class that knows what an agent is. Site is up at coconutos.org; the kernel fork and the distro tree are private.
Stage-analysis screener for Indian equities. SEPA rules applied to the NIFTY-500: VCP detection, IBD-weighted RS percentile, a seven level exit hierarchy, walk-forward backtesting, TradingView Pine export. Private while it bakes.
Built by @ShreyPatel4. Co-built with @jaypatel15406. Banner is generated by scripts/make_banner.py.