About
I build production ML systems. I’ve worked on real-time ad decisioning, applied research on acoustic-scene monitoring, and most recently agentic GenAI for clinical research at Gilead Sciences.
M.S. from UC San Diego, where I was co-advised by Peter Gerstoft and Yoav Freund. AWS ML Certified.
Currently working with
Experience
Senior Software Engineer, GenAI
UsefulBI Corporation
- Maintained and extended a multi-agent Text-to-SQL system at Gilead that lets clinical analysts query trial data in natural language. Added multimodal visual QA that regenerates code when rendered charts fail validation, plus self-correcting retry loops in the visualization agent.
- Rewrote the data onboarding pipeline, fixed a diversity-of-results issue in the real-world data (RWD) agent, and diagnosed intermittent frontend-backend disconnections (resolved with exponential-backoff reconnection).
- Owned the GenAI side of an internal medical writing tool for Clinical Study Report generation, which I scaled to support other regulatory document types. Tripled document length from 30 to 100 pages, improved table and image formatting, added per-paragraph citations, and built an automated template generation feature on top of the existing template-based flow.
- Added LLM-as-a-judge QC for generated content, per-section regeneration, and page-number filtering in OpenSearch retrieval.
- Led cross-functional delivery across engineering, data scientists, and pharma subject-matter experts. A lot of the work was translating requirements between these groups.
Machine Learning Engineer
Decide Technologies Inc.
- Shipped a 100× reduction in inference latency on a revenue-critical ad-decision engine, unblocking real-time bidding at production scale. Approach: JIT-compiled the inference path with Numba; rewrote the hot loop for cache efficiency.
- Owned the team-wide model evaluation pipeline: one reusable framework adopted across every ML project, replacing ad-hoc per-model scripts.
- Built a parallelised execution system using multiprocessing for 4×+ speedup in model evaluation.
- Designed and deployed four ad-selection models reaching top-quartile performance, integrated into containerised CI/CD pipelines with Docker and GitHub Actions.
Data Science Researcher
Scripps Institution of Oceanography, UC San Diego · co-advised by Peter Gerstoft & Yoav Freund
- Research under Peter Gerstoft and Yoav Freund on acoustic-scene monitoring: using ad hoc microphone arrays to localize sound sources in indoor environments.
- Built data pipelines for cleaning and storing sensor data, and authored a Python analysis library used across the research group.
- Applied Principal Component Analysis and regression models to recover source locations from raw audio. Migrated sensor data infrastructure from Google Drive to MySQL for 100× faster preprocessing.
- Work contributed to a peer-reviewed paper in the IEEE Internet of Things Journal (March 2022).
Graduate Teaching Assistant
University of California, San Diego
- TA for DSC 100: Introduction to Data Management (Winter & Spring 2021). Led sections on SQL, relational modeling, and data wrangling.
- Tutor for ECE 15: Intro to Programming in C (Fall 2020). Supported undergraduate students with fundamentals of systems programming.
- TA for PHYS 1AL: Mechanics Laboratory (Spring 2020).
Signal Processing Intern
Wavelet Group
- Developed a radar signal processing algorithm for calculating the distance, speed, and angle of approach of moving objects.
- Synthesized and validated the algorithm on an FPGA.
- Built a Python GUI for real-time visualization of approaching objects.
Selected Projects
Resume RAG Chatbot
Retrieval-augmented chatbot using OpenAI embeddings, FAISS indexing, and a FastAPI backend. Deployed with CORS hardening, health checks, and token/latency logging, optimized for cost within free-tier constraints.
Time Series Caption Generator
Neural captioning model using CNN + LSTM in PyTorch with a custom noun-weighted loss function. Received top grade and recognition for novelty and rigor.
Instacart Recommender System
Collaborative filtering system with TF-IDF weighting, popularity adjustments, and product-category enrichment for personalized item recommendations.
Writing
A Closer Look at Two Sum II
Exploring the Two Sum II problem and why the two-pointer approach is optimal. Uses matrix and graph analogies to show how a sorted array enables an O(n) solution instead of O(n²).
Read articleAudio Scene Monitoring Using Redundant Ad Hoc Microphone Array Networks
Peer-reviewed research from my time at UC San Diego. A method for localizing sound sources using ad hoc circular microphone arrays, without prior knowledge of array positions. Demonstrated on five IoT 8-microphone arrays in an office setting.
View on IEEE XploreEducation
Awards & Honors
Government of Maharashtra Scholarship for Higher Education in Foreign Countries
A competitive state-government scholarship supporting students from Maharashtra pursuing postgraduate studies abroad, awarded on the basis of academic merit.
Runner-Up: Best Outgoing Student
Recognized for overall academic excellence and contribution at the undergraduate level.
English Proficiency: Perfect Scores
Perfect scores on both major standardized English proficiency exams.