Dev Purohit

Extended an existing Streamlit damage detection app to support real-time webcam feeds and video file uploads, expanding capability beyond its original image-only pipeline Benchmarked EfficientNet, InceptionNet, ResNet, and IJEPA architectures — identified ResNet as the best model achieving 93% accuracy vs. prior baseline of 68% (↑25pp improvement) Optimized inference by converting the trained ResNet model to ONNX and deploying it with OpenVINO, significantly reducing model latency and improving real-time inference speed

Artificial Intelligence Intern at Tredence Inc. (2026-02 – 2026-08)

• Extended an existing Streamlit damage detection app to support real-time webcam feeds and video file uploads, expanding capability beyond its original image-only pipeline
• Benchmarked EfficientNet, InceptionNet, ResNet, and IJEPA architectures — identified ResNet as the best model achieving 93% accuracy vs. prior baseline of 68% (↑25pp improvement)
• Optimized inference by converting the trained ResNet model to ONNX and deploying it with OpenVINO, significantly reducing model latency and improving real-time inference speed
• Trained a YOLOv10-Nano object detection model to localize target objects and crop the detected region before passing it to the damage classifier, enabling object-specific damage prediction while eliminating unnecessary background information
• Integrated ScoreCAM++ to generate visual heatmaps highlighting damaged regions, making model predictions interpretable and explainable for end users
• Expanded the training dataset with multiple new object categories beyond parcel boxes, improving model generalizability across a wider range of items
• Built a fallback signature verification pipeline for a drug testing company, extracting handwritten names and initials from log tables to create an auditable record of drug test reviewers
• Designed a two-stage matching system: OCR-based text matching as primary method, and image similarity matching as fallback — ensuring high recall for regulatory compliance
• Designed and deployed a SharePoint team portal enabling members to showcase project portfolios, demo videos, and deliverables — improving internal knowledge sharing across the AI team
• Produced a comparative LLM evaluation deck benchmarking Gemini, Claude (Opus & Sonnet), OpenAI GPT, and Qwen across performance, cost, and enterprise fit — presented by management to company leadership for internal AI tooling decisions

B.Tech in Computer Science and Engineering (AI & ML Specialization) – SRM Institute of Science and Technology, Ghaziabad (2022-06 – 2026-05)

Senior Secondary – St. Mary's Sr. Sec. School, Haridwar (2021-04 – 2022-07)

Matriculation – St. Mary's Sr. Sec. School, Haridwar (2019-04 – 2020-07)