Backend engineer · AI infrastructure · DevOps · Mumbai, India

I build the libraries I needed while shipping AI products that couldn't lie to users.

4 npm packages (one published with build provenance) · a GitHub Marketplace Action · 7 open-source repos, each with a live demo and green CI.

flowchart TD
  H["Homesty.ai — live AI product · sole engineer"]

  subgraph reliability["AI reliability"]
    A["Anchor · provenance RAG"]
    T["Tripwire · stream guardrail"]
    G["Goldset · LLM eval Action"]
    Q["Quickdraw · streaming benchmark"]
  end

  subgraph platform["Backend platform"]
    AN["Anvil · webhook to BullMQ · npm + provenance"]
    S["Stackup · Kubernetes GitOps lab"]
    C["Codecraft · in-browser IDE"]
  end

  H -->|patterns extracted| A
  H -->|patterns extracted| T

• 🔨 Anvil — Idempotent webhook → BullMQ worker pipeline. HMAC-SHA256 constant-time verify, dedupe, backoff [1s, 5s, 30s, 5m], dead-letter replay. Ships with a Terraform module and a Helm chart. On npm with build provenance.
• 🪨 Anchor — Provenance-first RAG with cosine-floor refusal: returns refused: true instead of guessing when no source clears the floor. Live playground.
• ⚡ Goldset — Three-runner AI eval framework as a GitHub Action. PR comment diffs the delta, merge-blocks on regression. npm + Marketplace.
• 🚦 Tripwire — Mid-stream LLM safety. Token-by-token rule engine, sub-millisecond abort on rule trip. Ships an OpenAI-compatible sidecar proxy that streams through the guard. npm.
• 📊 Quickdraw — LLM streaming benchmark CLI. TTFT, tokens/sec, $/1K. Nightly bench against OpenAI + Anthropic. npm.
• ☸️ Stackup — Production-shape Kubernetes locally in 10 minutes: ArgoCD app-of-apps + Argo Rollouts canary + kube-prometheus-stack via one make up. Docs.
• 💻 Codecraft — In-browser IDE that boots a real Vite + React dev server in the tab via WebContainers. Editable Monaco + interactive xterm + IndexedDB snapshot cache. Live.

On npm: @ykstormsorg — anvil, goldset, quickdraw, tripwire.

Sole engineer on Homesty.ai — a live buyer-side real-estate AI on Next.js 15 + Postgres/pgvector + Prisma + GPT-4o + Claude. Refusal-first retrieval and a mid-stream guardrail were extracted from this work into Anchor and Tripwire.

Backend: TypeScript · Node 20 · Postgres + pgvector · Prisma · Redis · BullMQ AI: OpenAI · Anthropic · RAG · LLM streaming · prompt-injection defense Infra: Docker · Kubernetes (kind, ArgoCD, Argo Rollouts) · Terraform · Vercel Observability: Sentry · Prometheus · Grafana

Backend-platform / AI-infrastructure / DevOps roles — remote-first, Bangalore or Mumbai startups, YC seed-stage founding engineer, or contract work on RAG, streaming LLM, and queue/webhook reliability.

📍 lakshyaraj-dev.vercel.app · 📧 raolakshyaraj@gmail.com · 🐦 @ykstorm