Siddharth Harsh Raj siddharthharshraj

Siddharth Harsh Raj

Senior Full-Stack GenAI Engineer · Bengaluru, India (UTC +5:30) · open to senior remote roles

siddharth-dev.tech · LinkedIn · 𝕏 @siddharthraj29 · contact@siddharth-dev.tech

$ siddharth --whoami

  role        Senior full-stack GenAI engineer: agents, RAG, evals, and the product around them
  level       I set the architecture, the guardrails and the bar for "done"
  reach       my work runs in 70+ organisations serving 11M+ beneficiaries
  owns        problem → architecture → CI/CD → observability in production
  believes    an agent is only as good as its guardrails, and its guardrails are only as good as its tests

🔭 Where I'm headed

AI that is trustworthy by construction, built for people who can't afford a wrong answer: NGOs, frontline health workers and small teams running on thin margins. Three convictions shape everything I build:

1. Agents are software, so test them like software. Evals, validators and regression suites ship with every release. 2. AI is judged in the field, not in the demo. Low-end phones, patchy networks, many languages, real stakes. 3. Safety and cost are product features. Tool-gates, human review and hard spend caps are designed in from day one.

🧠 The work I'm proudest of

Avni Skills SDK

Open-source agentic SDK · core engineer · adopted org-wide

An agent that is safe by construction, not by hope:

🔁 Every turn is a git commit, and a validator runs after each one
🚧 Domain rules are MCP tool-gates the agent must call. A hook auto-reverts anything it does off-script
💸 A hard $5/session cost cap plus a thrash detector
🧪 463 unit tests + a real-LLM eval harness scoring agent quality release over release

Avni AI

Production RAG + LangGraph · core engineer

From an NGO's programme spec to a field-ready mobile app:

📄 Forms, validation rules and logic generated by RAG + LangGraph, then human-reviewed
⏱️ Implementation cut from 2–3 weeks to ~1 day
🌍 Ships to 70+ orgs and 11M+ beneficiaries
🏅 Part of a recognised Digital Public Good

Agentic Support Ops · sole author · private

The AI drafts, a human decides. It is read-only and draft-only by construction, with no write-DB connection anywhere in the workspace. Measured on real production tickets: −60% ticket resolution time and +40% team efficiency.

🛠️ Things I've built end to end

Project	What it is	The number that matters
QuickPe	Real-time fintech wallet: ACID transactions, RBAC, JWT, audit trails	2,190+ concurrent users at ~391 ms, −70% DB load
MeTube	Video streaming platform with real-time chat	−92% API calls via caching and debouncing, chat at <35 ms
OpenCam	Offline-first browser camera suite	MediaDevices API + IndexedDB, works across devices and browsers

🧰 My toolbox, by layer

  ┌─ AI & agents ── RAG · LangGraph · MCP tool-calling · evals & guardrails · Claude · GPT · Llama
  ├─ backend ────── Node.js · FastAPI · REST · WebSockets / SSE
  ├─ frontend ───── React · Next.js · Tailwind
  ├─ data ───────── PostgreSQL · pgvector · MongoDB · Redis
  ├─ cloud & ops ── AWS · GCP · Docker · CI/CD · New Relic · Vercel
  └─ languages ──── TypeScript · JavaScript · Python

🧭 How I build

Measure before and after. Every claim on this page has a number, and the number came from production.
Guardrails are code. Cost caps, tool-gates, validators and evals ship in the same PR as the feature.
Humans stay in the loop wherever a wrong answer costs someone something real.
Own it to production. The work isn't done until it's observable and someone is using it.

🔥 I build AI that ships, and stays shipped: effective, scalable, measured.

_{Also on GitHub: @siddharthr29}

Navigation Menu