Mihir Agarwal

Hello! Welcome to my personal archive. 👋

I am a Master’s student at Columbia University, researching memory-augmented transformers and generative modeling. My goal is to build systems that are not only theoretically sound but computationally efficient at scale.

My current work focuses on replacing RAG pipelines with memory layers that store knowledge directly inside transformer attention as learned key-value banks (advised by Prof. Micah Goldblum, Columbia) — removing the need for external retrieval entirely. I also built an evaluation harness for autonomous database agents, benchmarking frontier LLMs across concurrency control, access control, integrity rules, and query efficiency — dimensions most benchmarks ignore (ICML 2026, under review). Beyond that, I worked on generative modeling, developing neuro-symbolic frameworks that integrate Neural Point Processes with Diffusion Models to capture complex continuous-time dynamics.

Professional Background

I approach research with a strong engineering discipline, honed during my three years as a Software Engineer at Shell. There, I engineered a production RAG pipeline serving 90,000+ employees and built ML-driven test prioritization tools, reducing execution overhead by 35%.

My research journey began at ISEP (Paris), where I co-authored a Springer Journal paper on ML-based resource management for distributed systems (Fog Computing). I also have experience deploying computer vision models for industrial drone inspections at Vidrona (London).

What’s Next

I am actively looking for research internship and industry opportunities for Summer 2026.

I am always open to discussing technical challenges or potential collaborations.