Abhay Kansal

Incubated and led efforts in ML Engineering, MLOps, Computer Vision along with data engineering to develop, maintain and deploy at scale. Trained and deployed domain-specific small LLMs/VLMs using LoRA for efficient structured data extraction via adapter switching. Architected a scaled POC consuming STT models and generating customized responses via ChatGPT, enabling a comprehensive 360 degree customer view. Enabled multi-tooling support for the chat agent, integrating image and hybrid search capabilities using typesense & mivlus (vector databases). Built a Small Language Model (SLM) to extract entities from text at real-time throughput, efficient for custom tooling for agentic interaction. Led and expanded a cross-functional AI/ML practice spanning Data Engineering, Deep Learning, ML Engineering & MLOps. Developed and deployed fine-tuned STT models (Hindi, Kannada, Tamil, Telugu), achieving >100 RealTime Factor (RTF) and reducing latency. Optimized deep learning deployment workflows using Triton, TensorRT, ONNX, and mixed/half-precision, maximizing hardware utilization. Implemented ensemble model scheduling with zero memory copy, significantly improving inference performance. Led the migration of the observability stack from managed Google services to a self-managed Grafana/Alloy/Loki setup. FFMPEG with GPU enabled image decoding and encoding, delivering speed and reduced memory copy for enhanced performance. Spearheaded the Snowflake data architecture evolution, ensuring data privacy compliance and managing migrations/upgrades. Architected real-time data pipelines with StarRocks/Kafka for Data Science, powering cost-effective recommendations. Evaluated Delta Lake/Iceberg formats, enabling compute-storage separation and multi query engine support.

Led development and productionization of various computer vision models, including image classification and object detection tasks. Defined and designed the base architecture for the data pipeline, covering modeling, storage, and inference on premise. Implemented real-time video pipeline with latency less than 200 ms for on-premise solutions. FFMPEG with GPU gave a 20x speedup in the pipeline. Led efforts in data science and engineering to develop and deploy AI based solution on edge & cloud at scale.

Led efforts on data modeling using ML/DL methods and established ETL strategies to enable insightful reporting. Responsible to define strategy and architecture of new products that makes products responsive to varied enterprise product requirements. Developed in-house green field platforms/products focusing on delivering BI/DS at scale. Developed various DL/ML energy forecast and anomaly detection models for commercial customers. Built energy disaggregation models using a CNN-based model instead of standard RNNs, converting the problem to device state detection. Defined the strategy and architecture of new data products responsive to enterprise requirements. Developed in-house greenfield platforms/products focusing on delivering BI/Data Science at scale. Led development of Advanced Analytics BI platform with subsecond queries, handling more than a million data points in browser.

Analyzed and modeled behavior using NGO donor datasets and studied consumption patterns.

Conducted research in numerical simulation, ran modeling and simulation jobs for nuclear risk assessment studies using proprietary codes.

Developed and studied a coupling interface for multiphysics & multiscale simulation problems for nuclear reactor study.

Conducted numerous design projects with varied criticality as R&D engineer.