This is my personal portfolio.
I work on the parts of a system where the clock is the spec. Six orders of magnitude sit between the two numbers on that axis, and most of what I build lands somewhere along it. A risk check that has to clear before the order misses the market. A scheduler deciding which tenant gets the next slot on a GPU. A driver moving blocks off a wire without touching the heap.
Rust and C++ where latency is the product. Python and Go for the parts that have to still make sense a year later.
Boston. Looking for a full time systems or infrastructure role, and happy to relocate anywhere in the US.
Not stars. Every row is a measurement with a repo and a command behind it.
| Measurement | Result | Repo |
|---|---|---|
| Quiet-tenant TTFT p99 while a neighbor floods the same A100 | 1,585 ms → 61.5 ms, back to 1.14x solo | kvwarden |
Pre-trade risk gate, 7 checks, no_std Rust, zero heap allocation |
37 ns, 23M evals/sec, P99 42 ns | risk-hotpath-hft |
| Streaming replay of a generated trading day through that gate | 1.5M events/sec, 780K events in 0.5 s | risk-hotpath-hft |
| Kappa pipeline on real AWS, deployed end to end, measured, torn down | 45 assertions, every claim mapped to an evidence tier | realtime-analytics-004 |
|
Tenant-fair LLM inference on one GPU. Sits in front of vLLM or SGLang, rate-limits per tenant at admission, and keeps a quiet user fast while a noisy neighbor floods the same engine. Ten lines of YAML, no application code change. Fills the gap between Ollama, which is single user, and Dynamo or llm-d, which want a datacenter.
|
A 500-line Config hot-swaps mid-stream with no P99 delta. Every check is property-tested.
|
|
A real-time analytics rebuild for a Series-B martech company: design plus working artifacts that produce measured numbers. Deployed live on AWS, measured, then torn down. Ships an evidence ledger that maps every claim to a tier, so nothing in it is inflated.
|
One canonical Stage-0 built exactly as a 2016-era Notion would have shipped it, broken to the smallest grain: block CRUD over FastAPI, Automerge CRDT merge service, Kafka fan-out, Meilisearch, auth, billing, dashboards.
|
|
Agents that drive a real browser through a natural-language task, track the flow in Postgres, and store screenshots and DOM snapshots in MinIO. Exports the whole run as a dataset.
|
A hybrid management plane that connects many Kubernetes clusters to a central control plane and delivers platform extensions safely. Agent watches a CRD and reconciles, control plane owns registration and desired state.
|
More projects
| Project | What it is | Stack |
|---|---|---|
| Coconut OS | New product direction for Coconut Labs: a Linux distribution where agents are first-class kernel primitives. Landing page at coconutos.org | Linux, kernel |
| FastLane_NVMe | Storage offload stack that exposes NVMe semantics to guests over RDMA, with FPGA-ready hooks | Rust, VFIO-user |
| AMMIS | Reproducible C++20 workspace for latency-focused research: Conan, CMake, Google Benchmark, sanitizers | C++20, CMake, Conan |
| dream_team | A 31-agent engineering organization running as a daemon, with per-agent identity files and chat bridges | Python, agents |
| Data-Kitchen (Nova) | AI-native customer-VPC data platform. Four sprints shipped, then archived after an honest pre-customer review. Handoff doc included. | Python, Iceberg, FastAPI |
| solution_SnowConvertAI | SQL Server to Snowflake migration with a verification harness and a head-to-head against SnowConvert AI | T-SQL, Snowflake |
| Latent-Diffusion-Artbench | Latent diffusion model trained from scratch on the full ArtBench dataset | PyTorch |
Data-Kitchen is archived on purpose. I would rather show the review that stopped it than pretend it is still alive.
What I actually reach for
| Layer | Tools |
|---|---|
| Systems | Rust (no_std, VFIO-user, RDMA), C++20, CMake, Conan, Criterion, proptest |
| Inference | vLLM, SGLang, token-bucket and DRR admission, per-tenant TTFT histograms |
| Data | Postgres, Kafka, Redis, Iceberg, Parquet, Meilisearch, MinIO, Snowflake |
| Cloud | AWS (Kinesis, Lambda, KMS), Kubernetes, Terraform, Helm, Cloudflare Workers |
| Observability | Prometheus, Grafana, OpenTelemetry, k6 |
| Everyday | Python, Go, FastAPI, Playwright, Docker, GitHub Actions |
Recent pushes
| Repo | Last commit | When |
|---|---|---|
| coconutlabs-site | gallery: every unit links its public source repo | 2026-08-10 |
| kvwarden | chore: v0.1.6 dependency hygiene — cut runtime deps 144 MB → 33 MB,… | 2026-08-08 |
| .github | Add org banner, rewrite the profile | 2026-08-08 |
Reach me at shreypatel@coconutlabs.org or patelshrey77@gmail.com. Full time roles, contract work, or an argument about inference fairness. All three are welcome.