Soumy Sawant · Undergraduate AI Researcher · Manipal University Jaipur

I research AI evaluation, Information Retrieval, RAG, and reinforcement learning, with a focus on understanding when retrieval and LLM systems are actually reliable.

My recent research:

• 📄 Metric-as-Feature Leakage: A Feature-Side Failure Mode in Retriever Routing
  Accepted Poster — NewInML Workshop @ NeurIPS 2026
• 🔎 RouteLeak-Eval — a reproducible toolkit for auditing metric-derived leakage in retriever routing
• 🧪 Researching retrieval evaluation, adaptive retrieval, and LLM/agent reliability

Currently building:

• 🔬 RAG and retrieval evaluation systems
• 🤖 Agent architectures and evaluation harnesses
• 🧠 RL-based decision-making systems
• 🦾 AI architectures for humanoid robotics

Information Retrieval · RAG · AI/Agent Evaluation · LLM Reliability · Adaptive Retrieval · Retriever Routing · Reinforcement Learning · Agent Planning

• Retriever Routing & Evaluation — BM25 / Dense / Hybrid retrieval, routing, leakage auditing, benchmark design
• Efficient RAG — BM25, FAISS, RRF, local LLM inference, reproducible QA benchmarks
• Agent Evaluation — 80-case evaluation suite across multiple reasoning and planning tasks
• RL Systems — REINFORCE, PPO, A2C, DQN and related decision-making algorithms

Philosophy: If I can't implement it, I don't understand it.

I learn in public — documenting experiments, failures, and research insights.