Nov 2025 – Present · 6 mos · Bengaluru, Karnataka, India · On-site

Oct 2022 – Sep 2025 · 2 yrs 11 mos

• Worked in Model Development Team of Exit Vision & Item Recognition as a Service(IRAS) in Sam’s Club
• Scalable Recognition Workflow: Designed and implemented an automated end-to-end classification and verification workflow using EfficientNet with Sub-center ArcFace/MagFace losses and DOLG models , optimizing data preprocessing, training, and evaluation, resulting in 87% Precision and 92% Recall through metadata-driven image clustering.
• Active Learning: Designed and implemented optimal sampling logic in inference results based feedback loop for efficient high-quality data collection.. Processes >1M transactions daily in 15 hrs (30x time reduction) for 600 Clubs and samples <3% transactions with 70% lesser data to upload along with maintaining 98.2% coverage of new products with low recognition
• State-of-the-Art Performance: Conducted proactive investigations for performance improvements, contributing to the development of an industry-leading item recognition model with 97%
• Precision and 96% Recall. These outstanding results surpassed competitors in Item recognition for Scrubber Vision. Got the Walmart Excellence Award for the same.
• Scalable Evaluation Pipeline: Established an end-to-end evaluation pipeline for the IRAS
• classification model on ML Platform, optimizing it to process 500k images in under an hour,
• achieving a 10x reduction in evaluation time while expanding dataset capacity by 12x.
• Published white paper titled 'Metric Learning based Shelf Item Recognition on Images from Autonomous Robots' at the prestigious internal Walmart conference, Spark Tech Summit 2024, selected as one of 32 poster presentations from over 1300 abstracts and 775 full papers submitted.
• Won President’s Innovation Award of the Year(2025), Tech Disruptor Award(2025), Bravo Award (2025), Team Award( 2024), represented Exit Vision in Converge 2024 (The flagship annual retail-tech event hosted by Walmart Global Tech) & 5th International Conference on AI-ML Systems 2025.

Jul 2021 – Sep 2022 · 1 yr 2 mos

• Enhanced IRAS Item Classification Model: Contributed Significantly in the development and implementation of model enhancements, achieving a 35% reduction in False Positives and 50% reduction in False Negatives through several algorithmic tuning and feature engineering.
• Model Optimization: Optimized the runtime of the IRAS item classification model, resulting in
• a 10x reduction in model size and an 8x decrease in inference time, significantly boosting
• operational efficiency.
• Automated Keyword Management: Designed and deployed a daily auto-generation and
• verification pipeline for Negative keywords, resulting in a 15% improvement in Precision,
• optimizing the IRAS data processing workflow.