• Designed and productionized an ultra–low-latency vision–language model (VLM) powering multimodal autosuggest in Bing Visual Search, achieving sub-100 ms p95 latency through layer pruning and inference optimizations.
• Lead technical direction for new multimodal and recommendation features across Bing Visual Search and Related Searches, partnering with engineering and product to align model capabilities with business goals.

• Modernized the Related Searches recommendation stack across SERP and Image verticals with deep-learning–based recall and ranking, contributing to 10%+ higher clicks and up to 50% broader coverage, and enabling finer traffic shaping and faster experimentation.
• Innovated and experimented with various SLM Finetuning techniques including Pruning, DPO etc. Fine-tuned and deployed customized student transformer models for web-scale inference, efficiently processing several billion queries on a regular basis.
• Conducted research on prompt engineering best-practices, and synthetic data distillation using LLMs from GPT4 family.
• Developed and launched the first prompt-safety neural network for Bing Image Creator (DALLE 3), significantly reducing over-blocking and harmful content leakage while preserving user experience and creative freedom for millions of users.
• Collaborated with Responsible AI, policy, and product stakeholders to identify potential misuse of generative images and establish high-quality measurement sets and fine-grained guidelines for sensitive categories.

• Generated actionable insights from user sessions to identify user frustrations and ship features to improve search experience.
• Conducted opportunity analysis for ads revenue growth and monetization.

Advisor - Prof. Marios Savvides

• Designed robust Fine-Grain neural network models for Retail Product Image classification task, Plugs and Out-of-stocks Detection
• Read and implemented latest research techniques in the field to optimize the model performance
• Working with Long Tail distributions
• Analyzed and cleaned large noisy real world image datasets

This work culminated to a patent for innovative contributions to the field.

Advisor - Prof. Conrad Tucker

Training a Deepfake video detection architecture using CNN and Pulse rate estimation

Specialization - Machine Learning

Coursework:

• 10-601: Introduction to Machine Learning
• 11-777: Multimodal Machine Learning
• 11-751: Speech Recognition and Understanding
• 11-755: Machine Learning for Signal Processing
• 11-785: Introduction to Deep Learning
• 16-720: Computer Vision
• 18-739: Security and Fairness of Deep Learning
• 18-785: Data, Inference and Applied Machine Learning
• 18-797: Pattern Recognition Theory