Software Engineer & Independent Researcher
I build software at Varsity Brands and research AI systems: how they are evaluated, how they run, and how they hold up in the real world. My work spans LLM inference, AI evaluation and measurement, and AI for healthcare.
I am a software engineer and independent researcher based in Chicago. I hold an MS in Computer Science from the Illinois Institute of Technology (2025) and a BTech in Computer Science from Cochin University of Science and Technology (2023).
My research looks at AI systems from a few angles: how they are evaluated and measured, and how they are served efficiently. My published work examines how evaluation labels for AI systems behave across public datasets. I am currently exploring LLM inference in depth: KV cache, quantization, speculative decoding, and the kernels underneath.
Healthcare is where I started and where my domain expertise sits. At Illinois Tech I taught graduate courses in digital health informatics and deep learning, and my engineering work keeps circling back to systems that have to be right, not just fast.
Zenodo preprint, October 2026. 384 annotated examples across AgentErrorBench and Who&When, coded on a unified 13-class scheme.
International Journal of Advanced Computational Engineering and Networking, October 2022. Extractive summarization using MPNet embeddings clustered by self-organizing maps.
The serving layer: KV cache management, quantization, speculative decoding, batching, and GPU kernels. Where model quality meets systems engineering.
How AI systems are measured, benchmarked, and understood: evaluation methods, datasets, and what the numbers actually mean. Measurement first, opinions second.
Digital health informatics and applied deep learning in clinical settings, the area I taught at Illinois Tech and still call my home turf.
Hybrid.
Remote, Chicago.
Digital Health Informatics & AI.
Deep Learning.
Code and data supplement for the cross-dataset paper: 384 annotated examples from AgentErrorBench and Who&When on a frozen 13-class unified scheme, with the full analysis pipeline.
Private video Q&A that runs fully in the browser. Local inference with Transformers.js and ffmpeg.wasm: your files never leave your machine.
A small GPT built from first principles: custom byte-pair encoding tokenizer, Transformer decoder, trained on WikiText. Written up as a long-form explainer on the blog.
Extractive text summarization from my undergraduate research: sentence embeddings clustered with self-organizing maps to pick the sentences that carry an article.
How planning to begin quietly becomes a way to never begin.
Local AI, WebGPU, and why the most private inference is the kind that never leaves the device.
Memory, skills, and the difference between personal context and repeatable workflows.
A 21-minute deep dive that ends with a working miniature GPT.
More on the blog: blog.abhinavtb.com