kamals1999
Remote: Yes, also happy to be onsite
Willing to relocate: Yes, anywhere
Technologies: Python, PyTorch, Hugging Face, LLM agents, MCP, RAG, GraphRAG, LoRA/QLoRA, TransformerLens, FastAPI, React, TypeScript, Node.js, MySQL, Docker, AWS, GCP
Résumé/CV: https://sherawatkamal.github.io
Email: [email protected]
LinkedIn: https://linkedin.com/in/kamalsherawat
Can you tell from an LLM agent's hidden states that it is going to fail a task before it actually does? In my research, mostly yes. I probed five open source agents (Qwen and Llama, 7B to 72B) across 1,500 WebShop and 187 ALFWorld tasks, and success is linearly decodable mid task with AUROC up to 0.85. The surprising part: when I patched that signal in a live environment, behavior did not change. The agent knows it is failing, but that knowledge is not what drives its actions. Paper under review at AAAI 2027.
I also worked with an Amazon research team on 7k+ failed agent runs and raised success by 11.3% by targeting the moments where agents go wrong. If you're building with agents, this is the part that matters to you: I can debug them beyond reading logs and tweaking prompts.
Last summer I joined Faishion, a small SF startup, and owned their size recommendation model from training to production. It had to work for brand new users with almost no data, and it lifted customer satisfaction from 3.7 to 4.2. Then I shipped a Chrome extension that put those recommendations on third party retail sites, which grew cross platform engagement by 12.4%. Nobody handed me a spec for either; I scoped them, built them, and shipped them in under three months.
Before that I spent 3 years as the developer clients called when things were slow or outdated. I rebuilt a legacy desktop engineering tool as a React/Node web app now serving 1k+ concurrent users, and cut API latency from 1.2s to 180ms on a 50M+ row database.
Looking for an early engineer role at a startup building AI products, where I can own features from model to UI. MS at Virginia Tech, available January 2027. Authorized to work in the US.