// initialize

Lakshyaraj Singh Rao

Backend Engineer · AI Infrastructure · Full-Stack · DevOps

lakshyaraj@dev — zsh

lakshyaraj@dev:~$ ▋

lakshyaraj@dev:~$ whoami

Backend Engineer · AI Infrastructure · DevOps

lakshyaraj@dev:~$ cat focus.txt

Anvil — webhook→BullMQ pipeline

lakshyaraj@dev:~$ ls ~/projects

anchor tripwire goldset quickdraw stackup codecraft anvil

lakshyaraj@dev:~$ status

Open to backend / AI-infra roles

email github linkedin npm resume

Mumbai · Remote · Open to opportunities

// About

I build backend and AI-infrastructure systems and ship them to production. Right now I'm solo on homesty.ai, a commission-driven real-estate advisor running on Next.js, pgvector, and a refusal-first retrieval layer.

The reliability patterns I needed there became open source: Anchor (RAG that refuses instead of hallucinating), Tripwire (mid-stream LLM guardrails), and Anvil (an idempotent webhook → BullMQ pipeline with dead-letter replay). Same engine, made public.

I care about the unglamorous parts — constant-time HMAC verification, backoff schedules, golden-dataset evals that block a merge on regression. I'm open to backend / AI-infra roles and contract work on RAG, streaming LLMs, and queue/webhook reliability.

// Mission Status

Currently

Anvil v0.0.1 — webhook → BullMQ pipeline. 5 design contracts green in CI. HMAC-SHA256 constant-time verify, backoff [1s, 5s, 30s, 5m], dead-letter replay.

Shipped this week

›anvil v0.0.1 — webhook→BullMQ pipeline, 5 contracts green in CI
›tripwire v1.0.1 — mid-stream LLM guardrail
›quickdraw v1.0.1 — LLM streaming benchmark
›codecraft-ai live at codecraft-ai-tau.vercel.app

Next up

›anvil — examples/stripe + examples/github, publish @ykstormsorg/anvil to npm

Open to

›backend / AI infra roles at Mumbai or Bangalore startups
›YC seed founding-engineer roles
›contract work on RAG, streaming LLM, queue/webhook reliability

Location

Mumbai, India · Open to Bangalore / Remote

// System Architectures

~/homestyFlagship

Homesty.ai

Live commission-driven AI real-estate advisor.

Production AI advisor on Next.js 15 + pgvector + GPT-4o + Claude. Solo-built. Refusal-first retrieval and mid-stream guardrails were extracted from this work into Anchor and Tripwire — same engine, both public.

Next.js 15PostgrespgvectorPrisma 7GPT-4oClaude

live code

status● LIVE

rolesole engineer

stack6 systems

extractedanchor · tripwire

// production patterns, hardened into OSS

~/anchor● LIVE

Anchor

Provenance-first RAG that refuses to hallucinate.

Cosine-floor retrieval returns chunks when similarity is high, refuses when it isn't. Postgres + pgvector + Next.js. Live playground available.

Next.jsPostgrespgvectorPrisma

live playground code

~/codecraft-ai● LIVE

Codecraft AI

In-browser AI IDE. WebContainer-backed live coding playground.

Boots a real Vite + React dev server in the browser tab via WebContainers. Editable Monaco is wired to the WebContainer filesystem with debounced writes and sub-2s hot-reload; an xterm terminal runs real shell commands; an IndexedDB snapshot cache drops repeat-visit boot under 20s. COOP/COEP cross-origin isolation is what makes SharedArrayBuffer and the in-tab runtime work.

Next.jsWebContainerTypeScriptMonaco

live code

~/tripwire◆ NPM

Tripwire

Mid-stream LLM safety. Catch the lie before the user finishes reading it.

Pattern engine watches LLM streams token-by-token, aborts on rule trip. Sub-millisecond decision latency. Ships an OpenAI-compatible sidecar proxy (POST /v1/chat/completions) that streams responses through the guard and aborts mid-stream on a trip.

TypeScriptNodeStreams

code npm

~/goldset◆ NPM

Goldset

Golden datasets + LLM-as-judge + structural assertions, as a GitHub Action.

Three eval runners (golden dataset, LLM-as-judge with a pluggable OpenAI/Anthropic provider, structural assertions), packaged as a GitHub Action that runs on every PR, posts a delta-vs-base comment, and blocks the merge on regression.

TypeScriptGitHub ActionsOpenAI

code npm

~/quickdraw◆ NPM

Quickdraw

Benchmark LLM streaming. TTFT, TPS, $/1K tokens.

CLI runs prompt sets against models, captures TTFT + tokens/sec, diffs against baseline. Nightly bench wired.

TypeScriptNodeOpenAIAnthropic

code npm

~/stackup◇ REPO

Stackup

Production Kubernetes locally in 10 minutes. ArgoCD + Argo Rollouts + Grafana.

make up bootstraps a kind cluster with ArgoCD, Argo Rollouts, and Grafana/Prometheus. For learning real k8s without a cloud bill.

KubernetesArgoCDHelmkind

code

~/anvil◇ REPO

Anvil

Idempotent webhook → BullMQ pipeline. HMAC-SHA256, backoff, dead-letter replay.

Verify (constant-time HMAC) → dedupe → enqueue → retry [1s, 5s, 30s, 5m] → dead-letter → replay. pnpm workspace; 5 design contracts proven by tests in CI. v0.0.1.

TypeScriptExpressBullMQRedis

code

// Technical Arsenal

Languages

TypeScriptPythonSQLBashYAMLGo (learning)

Backend/Data

Node.jsNext.js 15PostgrespgvectorPrisma 7RedisBullMQ

Infra/AI

DockerKubernetesArgoCDHelmGPT-4oClaudeOllamavLLM

// Live Telemetry

@ykstormsorg/tripwire

…