Sarvesh Khetan

I’m an AI Researcher building agentic systems that improve reliability, efficiency, and long-term memory for real tasks. At Sentient Labs, I implemented automated agentic harness tracing (Harbor ATIF) for EvoSkill, designed SERA for crypto agents with embedding-based tool routing and parallel execution (90% behavioural similarity to ReAct with 50% lower inference latency and 30% fewer tool failures), and extended GEPA using multi-model ensembles with LLM-as-judge scoring to cut optimization time by 25%.

Previously, I turned research into production-grade pipelines: as a Software Engineering - AI Research Intern, I optimized a multi-agent ReAct-style Deep Research system with LangGraph (40% improvement in response quality), built a RAG-based Azure Blob connector for 1,000+ file ingestion, and created a long-context LLM evaluator with LLM-as-a-Judge (94% agreement with human labels). Earlier, as a Data Scientist - AI Engineer at Piramal Capital & Housing Finance, I led a team of 5 to design a patented RAG-based Text2SQL Graph AI agent that secured $1M+ in management funding, upgraded unimodal RAG to multimodal VLM RAG (+35% retrieval relevance), and deployed open-source LLMs via CI/CD on ECS using Docker (30% lower build latency). I bring this momentum to my research work too—such as building Evodata and TraceDB for cross-iteration memory recall, and publishing ROMA results with 10% SOTA improvement on SEAL-0—while continuing to develop strong ML fundamentals through my Machine Learning TA roles.