I build systems that are production-oriented, performance-driven, and reproducible — from low-latency C++ engines and paravirtual drivers to Python ML/automation tooling, tenant-fair LLM inference, and distributed control planes.

• 🔭 Current work: founding Coconut Labs — tenant-fair LLM inference orchestration (kvwarden), observability, and Generative AI engineering
• 🌱 Interested in: systems, distributed control planes, low-latency trading infra, tooling, and ML for system observability
• ⚡ Fun: game QA automation, synthetic data & logs, and hardware-accelerated I/O

• Founded Coconut Labs — independent inference research lab at coconutlabs.org. Shipped org profile, landing page (ccocnutlabs-LP), 4-dot brand mark, and contact / newsletter rebrand off the kvwarden domain.
• Launched kvwarden v0.1.1 → v0.1.3 — tenant-fair LLM inference orchestration on a single GPU (no Kubernetes). Renamed from InferGrid and ran a multi-gate fairness ladder on H100 / A100 with vLLM and SGLang covering Llama-3.1-70B (TP=4), Mixtral-8x7B MoE, and mixed prompt-length distributions. Highlights:
  • DRR-priority admission with per-tenant token-bucket rate limiting (closed a 523× starvation gap to baseline).
  • Per-tenant TTFT histograms, fairness Grafana dashboard, opt-in anonymous telemetry receiver on Cloudflare Workers.
  • Streaming router fixes (admission-slot lifetime, real TTFT, max-stream-duration fence), interactive CLI + doctor + man pages.
  • Show HN launch with one-pager, FAQ, pitch, and architecture overview.
• kvwarden — Gate 3 / T2 (cache-pressure admission) — RFC + KV-eviction config + runbook, T2 admission test skeleton, TenantPolicy + tenant_id surface stub, M4 bench harness flags (--prefix-overlap, --bias-flooder-cost), py3.13 CI, and a one-command Docker Compose eval bundle.
• Minierva-SEPA (private) — algorithmic SEPA (Specific Entry Point Analysis) for Indian markets: NIFTY-500 screener with VCP detection, IBD-weighted RS percentile, 7-level exit hierarchy, walk-forward backtesting, TradingView Pine export.
• mlxd (private) — research + strategy docs for tenant-fair LLM inference on Apple Silicon (pre-product).
• solution_SnowConvertAI_final — SQL Server → Snowflake migration take-home with a verification harness and head-to-head comparison vs. SnowConvert AI.
• SP-K8s-Control-Panel-Using-Streamlit — Streamlit-based Kubernetes control panel for deployment scaling and pod operations.
• dream_team — a 31-agent engineering organization running as a daemon, forking OpenClaw as the gateway/UI layer (Slack / Discord / Telegram bridges, web dashboard) with per-agent identity files (SOUL.md, AGENTS.md).
• risk-hotpath-hft — overlay analytics + reporting with audit/compliance, local-first with optional cloud switches.

• kvwarden — Tenant-fair LLM inference orchestration on a single GPU. No Kubernetes. https://github.com/coconut-labs/kvwarden

• UI-State-DOM-Capture-Multi-Agent — Modular, scalable Python-based UI state capture agent designed to automate and document user flows in web applications. https://github.com/ShreyPatel4/UI-State-DOM-Capture-Multi-Agent

• stage-0 — One canonical Stage-0 exactly as 2016-era Notion would have shipped, broken down into the smallest grain. https://github.com/ShreyPatel4/stage-0

• OpenPlay_Tester — Autonomous QA agent for a tiny Godot game with full data pipelines, RL/IL play, video anomaly detection, OOD checks, and an LLM triage assistant. https://github.com/ShreyPatel4/OpenPlay_Tester

• FastLane_NVMe — Paravirtual NVMe with RDMA and FPGA-ready hooks (Rust systems & drivers). https://github.com/ShreyPatel4/FastLane_NVMe

• Adaptive-Market-Microstructure-Intelligence-System-AMMIS — Real-time pipeline and ML engine that ingests live tick data and delivers deterministic-latency C++ signals. https://github.com/ShreyPatel4/Adaptive-Market-Microstructure-Intelligence-System-AMMIS-

• ArcBridge — Hybrid resource projection & extension platform for managing many Kubernetes clusters and projecting their state into a central control plane. https://github.com/ShreyPatel4/ArcBridge-Hybrid-Resource-Projection-and-Extension-Platform

• hyper-realistic-synthetic-logs-generator — Minimal, extensible framework to generate hyper-realistic synthetic logs and metrics with configurable characteristics. https://github.com/ShreyPatel4/hyper-realistic-synthetic-logs-generator

• risk-hotpath-hft — Overlay analytics and reporting with strong audit/compliance features; local-first with optional cloud switches. https://github.com/ShreyPatel4/risk-hotpath-hft

(Additional repos: Data-Kitchen — A brewing hot product idea)

• ShreyPatel4/Data-Kitchen

Nothing added to the journal yet.

RSS not configured.

5dd0ed1 - chore: update README (repo-of-week & journal) (by github-actions[bot])

Based on public repositories:

• Python, Rust, Go, C/C++, GDScript
• LLM inference & serving: vLLM, SGLang, tenant-fair scheduling (DRR + token-bucket), per-tenant TTFT/Grafana telemetry
• Systems engineering: low-latency C++ pipelines, Rust drivers, paravirtual devices
• Automation & tooling: Python agents, synthetic data, QA automation
• Cloud & infra: Kubernetes, observability, CI/CD, Cloudflare Workers

Open to collaborations, contract work, and speaking about systems & infra. Best contact: patelshrey77@gmail.com

This profile README is available under CC0 — reuse as you like.