Projects

Self-Healing Error Recovery System

Built continual-learning, run-time query rewriting systems that learn from interaction data, enabling on-the-fly correction for long-tail entities in speech recognition.

Language Understanding

Created deep-learning and generative models to break down user questions into component queries, enabling multi-domain reasoning.

Data-Efficient Learning

Scaled Alexa to use semi-supervised, weak-supervised, and active learning, cutting the need for extensive human annotation when training LLMs.

Search Relevance & RAG

Developed query context prediction models that reduced search latency by 1/8th, improved CTR by ~1% via relevance tuning, and built trending search algorithms with anomaly detection.

Contextual Advertising

Created ContX at Amazon Ads, an NLP-based recommendation engine leveraging CPD pricing to show relevant product ads for any webpage.

Spoken → Written / Written → Spoken

Created models that accurately convert spoken utterances to text and vice versa, enabling robust transcription and natural-sounding TTS.