Tanya Nair

ECE @ Johns Hopkins

Blog

Perspectives on emerging research, trends, and innovation in technology and beyond. Repurposed from my Substack, @tanyagents.
Don't Let an LLM Be Your Toxicologist: Building mutabench
Four leading LLMs tested on 30 compounds from the Ames mutagenicity dataset missed 27–40% of known mutagens — and most were confidently wrong. Part 1 of a series introducing mutabench, a benchmark for evaluating LLMs on toxicity screening. May 25, 2026.
Lost in Translation: LLMs, Low-Resource Languages, and the Digital Divide
How limited linguistic diversity in LLMs deepens the digital divide — the data scarcity, corporate incentive failures, and infrastructure gaps keeping low-resource language speakers out of the AI future, and what Meta, Masakhane, AI4Bharat, and others are doing about it. May 18, 2026.