Work Experience
Applied AI @ Comcast, Chennai, India
Machine Learning Engineer 3
Time: Jul 2021 - Jul 2024
Topics: Full-stack (end-to-end) ML, Software Engineering (Backend), Data Engineering, Software Architecture Design
- Designed and developed AI for Operations project backend in microservices-based architecture hosted on AWS using Kubernetes with Airflow DAGs and real-time inference using state-of-the-art (SOTA) time-series models like Anomaly Transformers, TimesNet, N-HiTS, N-BEATS and Prophet.
- Developed auto-scalable APIs using FastAPI and handled production server deployment with a scale of over 1 million calls per day at an optimal point considering performance and cost.
- Built the event-driven architecture for proactive anomaly alerting and system dependency graph-based Root Cause Analysis (RCA) using Dynamic Time Warping (DTW) distance.
- Led the R&D of the log data mining pipeline using the Drain3 algorithm integrated with features like log trend anomaly alerting, log RCA and Q&A triage bot using Large Language Models (LLMs).
Data Science @ Jupiter, Bengaluru, India
Data Science Intern
Time: Jul 2020 - Dec 2020
Topics: Natural Language Processing, Deep Learning, Artificial Intelligence, Statistics, Software Development
- Designed data structures for Redis cache in Voice Annotation Platform, crucial for training Speech Models on Indian accents. Developed a generic Singleton Class module, for broader caching usage.
- Performed the research work on the SMS NER project using the Flair model and improved on the existing regex-based approach by 20% and further reduced inference latency by 75% from 1.2 seconds to 300ms per text using embedding optimization and quantization.
- Worked on advanced SQL and Airflow using Spark engine on Big Data for early data drifts.
Data Storage (ASIC) @ Sandisk (Western Digital), Bengaluru, India
Analog Design Intern
Time: July 2020 - August 2020
Topics: Analog Electronics, Electronic Devices, Microelectronics Design
- Designed a Band Gap Reference (BGR) Voltage Source in Cadence and making it robust to PVT variations.