Narasimha Karthik J
AI / Research Engineer — Evals, Post-Training, RL Environments
Hi, I’m Karthik. I build and evaluate AI systems: RL environments and graders for agents, post-training experiments, and LLM-as-judge pipelines that ship to production.
Open to AI Engineer and Research Engineer roles — evals, post-training, RL environments, agent systems. Based in Chicago, open to SF and NYC. Reach me at narasimhajwalapuram2026@u.northwestern.edu.
Recent Work #
Customer-Service Agent RL Environment — A τ²-bench-based environment where an LLM customer with a hidden goal talks to a tool-using agent. The reward gates on the database diff before an LLM judge checks consent and honesty. Red-teaming found the judge accepting forged “CUSTOMER: Yes” lines; speaker-tagged messages fixed it.
Sycophancy Recovery Study — I fine-tuned sycophancy into Qwen3-8B and compared five ways to remove it. DPO looked fixed to a 72B judge, but a probe still found the disposition and it relearned in 5 steps. Constitutional-AI preference pairs did best. Six-part write-up.
LEAR and QD-LEAR — LLM-driven evolution of agent-based rules (GECCO ‘25, ALife 2025), at Northwestern’s CCL Lab.
Experience #
WhoBlinked — Co-Founder & CTO (2026–present). Competitor-intelligence SaaS; LLM features checked against live search before users see them.
CCL Lab, Northwestern — Research Engineer (2024–present). Built NetLogo’s official multi-provider LLM extension and the code generator behind LEAR.
Elevate Packaging — Software Developer, contract (2026). Rollstock traceability system on FastAPI and Cloud SQL.
Relativity — Applied Science Intern (2025). LLM-as-judge pipeline over 1M+ legal documents; model routing cut review time 60% and cost 40%. Built a 600-example red-team benchmark that became the default release gate.
Boeing Research & Technology — SDE-2 & Data Scientist (2022–2024). RAG over 20k+ aircraft manuals that cut manual authoring effort 80%; grew from a solo prototype to $200k funding and a 6-engineer team.
Writing #
Technical posts here and at NeuraForge — post-training, RL, and transformer internals.