Experience - Charith Purushotham

Owl AI, Boulder, CO

September 2025 - Present

Developing a sports analytics pipeline leveraging fine-tuned YOLO, HRNet, RTMPose, and 3D pose estimation models for motion understanding.
Engineering GPU-optimized inference pipelines with quantization (TensorRT), caching, and prefetching techniques, reducing end-to-end latency by 50% (18→9s).
Curating domain-specific sports datasets and employing GenAI agents (Roo, Claude) for automated data labeling, validation, and preprocessing.

Python PyTorch YOLO HRNet RTMPose TensorRT GenAI Computer Vision

Idaho National Laboratory, Idaho Falls, ID

June 2025 - August 2025

Contributed to an LDRD-funded project developing and fine-tuning deep learning models (SAM/SAM2, YOLOv9, ResUNet, DenseNet, ResDense-UNet) for semantic segmentation of pores and cracks in post-irradiated nuclear materials using FIB, SEM, TEM, and X-ray CT data.
Utilized HPC infrastructure for large-scale training and image characterization, enabling predictive modeling of radiation-induced material degradation.
Analyzed how data quality factors (resolution, contrast, noise, artifacts) affect segmentation accuracy and applied saliency maps with attention mechanisms (CBAM vs. bottleneck self-attention) across microscopy modalities.
Supported a parallel study on size effects in nuclear materials, leveraging Random Forest and XGBoost to compare temperature sensitivity and impact energy between sub-sized and full-sized specimens.

Co-authored: "Bridging Multimodal Microscopy for Advanced Characterization on Nuclear Fuel Using Machine Learning", published in Frontiers in Mechanical Engineering – Digital Manufacturing, Vol. 11, 2025, demonstrating a transfer-learning deep-learning framework outperforming four state-of-the-art models for cross-scale defect segmentation.

View Publication

Python PyTorch SAM/SAM2 YOLOv9 ResUNet DenseNet Random Forest XGBoost HPC

Phamax, Bengaluru

September 2022 - July 2024

Collaborated on the system architecture for the digital assistant "Ariya," now a Microsoft-certified SaaS on the MS Teams store. Contributed to design, boosting subscriptions by 5% and client engagement by 7%, demonstrated by a higher client demo rate.
Implemented GPT features with OpenAI and Azure, enhancing data summarization and insights. Dockerized the solution, resulting in a 25% resource efficiency increase and reducing deployment time by 50% (from 15 minutes to 7 minutes).
Managed user and group administration on Azure, deployed Docker containers for efficient application deployment, and implemented automated logging and monitoring systems for improved user experience.

Collaborated on the content processing system architecture, which handles data from the client's IRMS to the SFTP data container.
Implemented a custom Q&A system for user queries and an application logging system to record conversation transcripts. The logging system increased user engagement with the chatbot by 5% due to regular feedback from monitoring the transcripts.

Led a data engineering project for Santen using Databricks and Azure, halving processing time from 30 to 12-15 minutes and improving data accuracy by implementing checkpoints for data evaluation.
Implemented a medallion architecture, enhancing Tableau performance by offloading heavy processing to Databricks.

Designed a Retrieval-Augmented Generation (RAG) pipeline using AWS Bedrock for pharma document summarization.
Deployed an Adverse Drug Reaction (ADR) detection model on Amazon SageMaker for enhanced drug safety monitoring.

Node.js React.js Bot Framework Flask Azure Docker Databricks AWS Bedrock SageMaker

Fittlyf, Bengaluru

April 2022 - August 2022

Developed a Multi-Arm Bandit framework for an in-house reporting tool, optimizing A/B test traffic and reducing turnaround time by 40%. Implemented Group Sequential Sampling to adjust significance thresholds, accelerating decision-making and cutting turnaround time by 46%. Automated logging of metrics and report generation enhanced overall efficiency.

Developed an MVP AI-based solution in the computer vision domain, using human pose estimation (Blazepose model) and deep learning, to help physiotherapists monitor and report the status of exercise routines for their patients.

Python Jupyter OpenCV Flask Tableau Excel BigQuery

Trawello Healthcare, Bengaluru

March 2021 - October 2021

Developed a streamlit-based Reporting Framework, which dealt with the Startup's data, which majorly consisted of student healthcare data. This Framework involved data analysis, data visualization, and ML prediction.

Python Streamlit ML Analysis

Experience Timeline