Backend Engineer · AI Infrastructure · DevOps — Mumbai, India
I build the reliability layer that lets AI run in production without supervision — refusal when retrieval fails, an abort when output goes off-script, eval-gating before merge, idempotency under webhook re-delivery.
Four npm packages (one with SLSA build provenance) · a GitHub Marketplace Action · seven open-source repos, each with a live demo and green CI — every one extracted from a live, solo-built AI product.
Diagram source (Mermaid)
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
| Project | What it solves | Links |
|---|---|---|
| Anvil | A re-delivered Stripe/GitHub webhook double-fires your worker. Anvil is the idempotent webhook → BullMQ pipeline that dedupes, retries on a fixed backoff, and dead-letters with replay. Constant-time HMAC verify. Terraform module + Helm chart. | repo · npm · SLSA provenance |
| Anchor | RAG that refuses instead of hallucinating — returns refused: true when nothing clears the cosine floor, with a provenance trail on what it does return. |
repo · playground |
| Tripwire | Mid-stream LLM guardrail that kills the response on a rule trip before the bad token reaches the user. Drop in as an OpenAI-compatible proxy (one URL) or a library. | repo · npm |
| Goldset | CI for AI — golden-dataset + LLM-judge + structural eval runners as a GitHub Action that posts a delta-vs-base PR comment and blocks the merge on regression. | repo · marketplace · npm |
| Quickdraw | LLM streaming benchmark CLI — TTFT, tokens/sec, p50/p95/p99, and cost per 1K across OpenAI + Anthropic, with quickdraw diff for regressions. |
repo · npm · SLSA provenance |
| Stackup | Production-shape Kubernetes on your laptop — ArgoCD app-of-apps + Argo Rollouts canary with a real Prometheus success-rate gate + kube-prometheus-stack, from one make up. |
repo · docs |
| Codecraft | In-browser IDE that boots a real Vite + React dev server in the tab via WebContainers — editable Monaco, live npm install in an xterm, snapshot-cached boots. |
repo · live |
On npm: @ykstormsorg — anvil, tripwire, goldset, quickdraw.
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 hardened out of this work into Anchor and Tripwire — same engine, made public.
Backend TypeScript · Node 20 · Postgres + pgvector · Prisma · Redis · BullMQ AI OpenAI · Anthropic · RAG · LLM streaming · prompt-injection defense · evals Infra Docker · Kubernetes (kind, ArgoCD, Argo Rollouts) · Helm · 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.