Amogh Gupta

Aligned models frequently over-refuse — they decline a sensitive prompt outright, or return generic safety boilerplate, when the user's underlying information need could have been met safely. SHARD introduces self-reframing distillation as a way to recover that lost helpfulness without weakening safety. In parallel, I audit what alignment leaves behind: my FAccT 2026 paper shows how bias in LLM-generated educational explanations compounds across language, caste, and social context rather than simply adding up across Indian and American STEM settings.

At Amazon, on the Alexa Audio team, I built A3LS, an accent-audio testing service that synthesizes speech across 13 accents and measures how much speech-recognition accuracy degrades for each one — a fairness question wearing the clothes of an infrastructure problem. Before that I worked on budgeting and financial planning software at Nomura, built a retrieval pipeline over 10,000+ documents at Ernst & Young, prototyped voice and agentic AI systems at an early-stage startup, deployed recommendation algorithms at Jio Platforms, and ran anomaly detection across more than a million logs at DRDO. What ties these together is that shipping something forces you to say precisely what "working" means, and that discipline transfers to alignment research more directly than I expected.

I completed my B.Tech in Computer Science at the Indian Institute of Information Technology Pune, graduating with Honors in Machine Learning ranked first in my batch, and placing consistently among the top three of 175 students. During my undergraduate years I was also a research intern at IIT Kharagpur, working on red-teaming and multilingual jailbreaks, and at IIT Indore on disaster-tweet summarization — the projects that first pulled me toward safety and evaluation as the questions I wanted to spend time on.