Reetu Raj Harsh

About Reetu Raj Harsh

AI/ML Engineer focused on Generative AI, Knowledge Graphs, and Responsible AI - building systems that are not just powerful, but trustworthy and production-ready.At Domyn, I\'ve gone from designing multi-agent financial advisors to deploying LLM models across 16 H100s on supercomputing clusters. I architect end-to-end AI systems: from GraphRAG pipelines on 20M+ financial relations to synthetic data generation at 1.3M+ LLM calls with automated quality evaluation.What drives my work:• Scale - I\'ve benchmarked graph databases, optimized tensor/expert parallelism, and run sharded parallel workloads on HPC infrastructure.• Responsibility - I built AI-Sentinel, a production Responsible AI platform with modular guardrails and full observability, because safe AI isn\'t optional.• Research - My work has been published at VIS 2025, accepted at NeurIPS Workshop 2025, and is under review at KDD 2026 and ICLR 2026.Previously at Tata Consultancy Services, I delivered GenAI solutions across healthcare and enterprise - from NeMo Guardrails-powered chatbots to agentic RAG systems for automated grant writing.Awards: Microsoft Hackathon (Most Scalable Solution)| UC Berkeley LLM Agents Hackathon WinnerI hold a B.Tech in Computer Science (CGPA: 8.83) from Heritage Institute of Technology, Kolkata.If you\'re working on LLMs, knowledge graphs, multi-agent systems, or responsible AI - I\'d love to connect.

Experience

1. Junior Ai Engineer

   Domyn

   Jun 2025 — Present · IN

   AI Engineer at Domyn, a deep-tech company building Responsible LLM solutions and AI agents for highly regulated industries - working with in-house models Domyn Small and Domyn Large, deployed on the Domyn Agentic Platform.Knowledge Graphs & GraphRAGBuilt FinReflectKG, the largest temporal financial knowledge graph spanning 10 years of S&P 500 10-K filings with 20M+ relations - covering entity deduplication, normalization, and canonicalizationArchitected a hybrid GraphRAG pipeline combining Neo4j/Memgraph graph traversal, ChromaDB vector search, and BM25 with cross-encoder rerankingBenchmarked Neo4j vs ArangoDB with custom text-to-Cypher/AQL generation for multi-hop traversal at scaleResponsible AIBuilt AI-Sentinel - a modular Python SDK and async FastAPI microservice with production guardrails (PII, toxicity, prompt injection, hallucination detection) and Langfuse observabilityBenchmarked 14 open-source safety guard models (110M–20B params) on a custom 79K-sample NIST AI RMF-aligned dataset across 8 safety categoriesWon Most Scalable Solution at Microsoft Hackathon for an automated red teaming framework targeting LLM vulnerabilities in finance and healthcareDid benchmarking of 14+ guard models and developed a custom benchmark.LLM Infrastructure & ScaleDeployed multi-node multi-GPU vLLM serving (Qwen3-235B across 16 H100s) on MareNostrum 5 HPC using Ray-orchestrated clusters, tensor/expert parallelism, and NCCL over InfiniBandBuilt a synthetic data generation pipeline across 100 parallel shards with Evol-Instruct diversity and LLM-as-Judge quality evaluationPerformed full SFT of Qwen3-0.6B as a financial safety guard model using TRL on distributed H100 clustersMulti-Agent SystemsDeveloped FinSage a production multi-agent AI Investment reaserch copilot with 12 specialized subagents across 78+ tools, real-time SSE streaming, Neo4j KG integration, NLP-based backtesting, real time sentiment analysis, risk assesment and Alpaca broker integration

Education

• Heritage Institute of Technology

  Bachelor of Technology - BTech, Computer Science and engineering

  2019 — 2023

• BR. DAV Public School

  12th(Intermediate), (PCM+IP)

  2016 — 2018