Jan 2026 – Jun 2026

Stateful Streaming ASR for Voice Agents

University of Washington · Advisor Prof. Mari Ostendorf

Inference-time state-management strategies for streaming ASR that preserve cross-utterance context while staying robust to long silences and backchannels. Roughly 20% relative reduction in turn-onset WER for FastConformer and Nemotron streaming models on Switchboard and CallHome, at sub-250 ms latency and under 0.6% memory overhead.

Speech RecognitionVoice AgentsStreaming
Sep 2025 – Mar 2026

Accent Bias in Audio Language Models

University of Washington · with Prof. Noah A. Smith

An evaluation framework for measuring accent-driven disparities in audio language models across hiring, academic, and English-test settings, using 700+ synthetic interview dialogues and a 68-participant human study. Received an Outstanding Paper Award at EMNLP 2026.

Audio LMsFairnessEvaluation
Nov 2023 – Oct 2024

ASR for Low-resource Indian Languages

Amazon × IIT Bombay · Advisor Prof. Preethi Jyothi

Parameter-efficient finetuning and text-only adaptation with cross-lingual transfer for multimodal multilingual ASR across 10+ Indian languages, including a 17% zero-shot WER reduction on IndicVoices-Odia.

Multilingual ASRPEFTSeamlessM4T
Summer 2024

Context-aware Time Series Forecasting

UT Austin · Advisor Prof. Sandeep Chinchali

A plug-and-play framework for injecting multimodal context such as news, market trades, and environmental signals into transformer forecasters, reducing MSE by 13.9% for PatchTST and 15.5% for iTransformer.

Multimodal LearningTransformersForecasting
Nov 2024 – Jun 2025

Hierarchical Inference for Edge AI

IIT Bombay · Master's thesis, Prof. Sharayu Moharir

HI-LCB, a UCB-style policy for deciding when a tiny on-device model should offload inference to a remote model, with order-optimal O(log T) regret and O(1) per-sample complexity.

BanditsEdge MLTheory
May 2024 – Oct 2024

Adversarial Attacks on Online Learning to Rank

IIT Bombay · Master's thesis, Prof. Sharayu Moharir

Observation-free attack strategies that promote target items in bandit-based rankers under cascade and position-based click models, achieving linear regret with logarithmic attack cost.

Online LearningAdversarial MLRanking