Om Bharatiya
Principal AI Engineer · Agentic Systems, Evals & MCP
📍 India
I build AI systems for places where a wrong answer costs money.
I work on the full lifecycle of production agents for an enterprise platform used by 150+ banks: the multi-agent harness, the eval infrastructure around it, and the monitoring that keeps it honest.
8 years across fintech, e-commerce, and enterprise AI · ~8k stars on open source · writing read by 80k+ engineers
Building
| Project | What it is |
|---|---|
| Litmuz | Claim-level verification for life-sciences research agents. Splits an AI-written memo into atomic claims, resolves citations against PubMed, PMC, and Crossref by rule, and returns a traffic-light verdict per claim. Safety-critical claims always route to a human. Web app and MCP server, so agents can check their own work. |
| AI Evals | A citation-grounded guide to LLM evaluation: error analysis, LLM-as-judge, RAG and agent evals, production monitoring. 26 task-type playbooks, runnable recipes, and free role-based courses. The public counterpart to the eval infrastructure I build at work. |
| AI Daddy | AI system design and interview prep for engineers moving into AI from backend, frontend, or data. RAG, agents, MCP, evals, and the production patterns behind them. |
| Searchcoins AI | Early-stage consumer AI for India's less affluent class: earn money, brand deals, and offers by learning practical things about your work, your money, and your rights. Voice or text, Hindi or Hinglish. |
Open source
| Repo | ★ | What |
|---|---|---|
| AI System Design Guide | 2.3k | How to design production AI systems and evals. Actively maintained; this is where my current thinking lives. |
| FAANG Coding Interview Questions | 5.5k | The curated interview prep list, forked 1.3k+ times. |
| TranscriptAI | - | Multilingual speech-to-text CLI on OpenAI Whisper. Batch processing, 8+ formats. |
| Screwdriver CI/CD | - | 10+ merged PRs across UI, models, and core. Shipped GitHub/GitLab PR validations used by 10k+ engineers. |
Writing · The Real Signal
Field notes from production agentic AI, plus sharper takes on the industry around it:
- How to Evaluate LLMs Without a Gold Dataset
- How Claude Manages Memory Across Context Windows
- The Open-Weight Illusion: US AI Gating vs China's Open Models
- Review Culture in India: How 5 Stars Became the New Tipping
Background
| Now | Principal AI Engineer at an Amsterdam-based banking AI company: multi-agent orchestration on LangGraph, MCP tooling, eval gates in CI/CD, production observability |
| Before | Founding Engineer at BRND.ME, India's fastest unicorn, now IPO-bound: led a 12-engineer team · document intelligence at 93%+ accuracy · agentic workflows automating finance and category operations · systems handling 10M+ requests/hr · a data orchestrator moving TBs of data a day |
| Earlier | Swiggy · HashedIn · AI/ML research intern at ISRO |
Stack: Java & Spring Boot · Python · LangGraph & LangChain · MCP · AWS (Lambda, SQS, Fargate, RDS) · Kubernetes · Next.js