ai + data engineer · coconut labs
I measure before I claim.
Then I publish what failed.
I build data systems and inference infrastructure, and I run a two-person research lab on the side. The through-line is not a stack, it is a habit: pre-commit the threshold, rent the hardware, run the bench, and keep the number that survived next to the caveat that limits it. Half my best results are disconfirms, and they are published like everything else.
26× tail recovery
6 pkgs + releases
7 in browser
3 and counting
how i think
Four habits, each with a receipt
Falsify before shipping
The first kvwarden thesis was an admission cap that would smooth the latency cliff. I rented an H100, ran 16,000 requests per arm, and the cap measured 1.04x, no effect. At one operating point it made things worse. The disconfirm is in the repo next to the result that replaced it.
gate 1.5 · 16k req/arm · published as-is
Caveats travel with numbers
The hero number is a 26x tail recovery for a quiet tenant under flood. The same sentence always carries the regime: one A100 at saturation, vLLM 0.19.1, n=311. On an H100 at the same load the starvation never happens, and the page says so.
53.9 / 61.5 / 1,585 ms · A100 · n=311
Names must be true
kvwarden's cache scaffold turned out to be a ledger nothing read. Three greps proved it, so I rewrote the roadmap in public, closed my own RFC as superseded, and put "the name says KV warden but you are not touching the KV cache" in the FAQ myself.
rfc closed as superseded · faq entry shipped
Stop digging when the hole is the tool
A benchmark provider burned 512 pod spins over four days without once handing me a working machine. I stopped, wrote up the $47 of evidence that the allocator was the problem, and moved the experiment rather than the budget.
512 pods · 0 usable · finding published
timeline · 2026
Showcase, in the order it happened
- aug 2026
One lab, one design system, everything live
Shipped the measurement-first redesign of coconutlabs.org, the Systems That Don't Lie masterclass, a public benchmarks proof page, and opened the source of five browser-runnable guardrail demos.
- aug 2026
kvwarden v0.1.6 on PyPI
Maintenance release that cut a clean install from 144 MB to 33 MB by removing dependencies the package never imported, and cleared four advisories CI had been installing for months.
- aug 2026
syn1, an audio unit from scratch
A C++ wavetable synth for Logic with no framework dependency in the DSP core: 330 test cases run in about ten seconds, aliasing measured at -64 to -118 dB, 2.4 percent of one core for 128 oscillator streams.
private repo · write-up: a model in the room
- jun 2026
Real-time analytics pipeline, measured on AWS
A take-home built as a production argument: seven runnable artifacts with measured throughput and cost numbers instead of a slide deck.
private repo, walkthrough on request
- may 2026
The 512-pod negative result
Four days of provisioning attempts across five GPU SKUs produced zero usable machines. Documented the allocator failure with logs and spend, then moved the bench to a different provider.
- apr 2026
kvwarden: fairness measured, name audited
Gate 2 established the tenant-fairness result on a rented A100. The same month a trademark audit found a senior user of the old name, so the project renamed in public with the reasoning attached.
- feb 2026
The lab starts
First commit of the inference middleware that became kvwarden. Two engineers, rented GPUs, a rule that every claim carries its bench.
Earlier work lives on the GitHub profile: an RDMA NVMe offload stack, a C++20 latency research workspace, a latent diffusion model trained from scratch, and an archived data platform kept public because the review that stopped it is worth more than pretending it lives.
working demos
Run them in your browser
| demo | what it proves | status |
|---|---|---|
| Silent data-regression guardrail | Catches schema-quiet corruptions a type check cannot see, on a profile fitted from clean episodes. | ● live |
| Point-in-time correctness | The one bug where offline accuracy goes up. Accuracy rewards it, schema checks miss it, this flags it. | ● live |
| Silent cache-miss guardrail | Feature-store staleness that never throws, made visible. | ● live |
| Columnar scan-bytes guardrail | Queries that quietly stop pruning, caught by their bytes. | ● live |
| Ingestion data-contract guardrail | Contract checks at the door instead of in the warehouse. | ● live |
| Agentic MLOps atlas | A full lifecycle map with 25 interactive stations. | ● live |
| The proof page | Every published number with its hardware, n, and raw artifact. | ● live |
Every demo runs client-side and links its source on GitHub. Recorded walkthroughs are planned; until they exist this page will not pretend otherwise.
learning · academic
Study one level below the job
The practice behind the shipped work is academic on purpose: a private systems atlas that goes from the business problem down to the metal, courses run against real prototypes, and a fifteen-minute rep every day. The public edition of that practice is the masterclass.
Systems That Don't Lie
A five-chapter masterclass on honest systems: tenant fairness, single-machine inference, nanosecond signals, agents from the ground up, and observability that cannot flatter itself.
The Library
The private academic wing: the waterline atlas and the study shelf. What earns publication is rewritten clean and lands as a research note.
notes
Written like lab records, not posts
| note | one line |
|---|---|
| Tenant fairness on shared inference | The measurement that started the lab's public life. |
| A model in the room | Three jobs a model has in a creative tool, and the fence that keeps it useful. |
| What mixing taught me about evals | Studio monitors and eval suites fail the same way. |
contact
Full-time roles, contract work, or an argument about inference fairness
All three are welcome.
[email protected] · [email protected] · github/ShreyPatel4 · github/coconut-labs