Sai Krishna
Bashetty
Architecting the future of Frictionless Retail and Safe Autonomy with scalable Computer Vision & Edge AI.
About Me
I am an Applied Science Manager and Computer Vision Architect with over 6 years of experience pioneering AI-driven Retail Automation and Safe Autonomous Systems. Currently at RadiusAI, I spearhead the research and development of ShopAssist™ and ShopAssist Pulse, industry-leading solutions for frictionless checkout and real-time retail analytics.
My work focuses on bridging the gap between state-of-the-art research and mission-critical commercial products. I have successfully architected edge-optimized vision systems that handle 10,000+ SKUs in real-time with 99.99% uptime, drastically reducing shrink and enhancing operational efficiency for major enterprise retailers.
With a Master's in Computer Engineering from Arizona State University, my background is rooted in rigorous innovation—from developing "DeepCrashTest" for autonomous vehicle safety (published in IEEE ICRA) to inventing novel multi-camera tracking and auto-calibration algorithms. I hold extensive patents in visual tracking, automated item recognition, and augmented reality for retail.
Technical Expertise
Professional Experience
Leading the Applied Science research and implementation team for ShopAssist™, an AI-powered frictionless checkout solution adopted by Fortune 500 retailers. Beyond my core science role, I assumed end-to-end technical leadership during the critical 0-to-1 (MVP) phase, successfully scaling the system from a pilot prototype to a revenue-generating commercial product.
Strategic Leadership & Product Delivery
- 0-to-1 MVP to Enterprise Scale: Led the end-to-end architectural design and development of the checkout system from its initial Request for Proposal (RFP) and MVP stages to a fully viable product deployed across hundreds of stores.
- Cross-Functional Orchestration: Directed collaboration between Applied Science, Engineering, Product, and Operations teams to translate ambiguous business goals into concrete technical roadmaps and deliver on aggressive timelines.
AI Architecture & Critical Problem Solving
- Solving the Data Bottleneck: Pioneered an unsupervised Computer Vision workflow to automatically annotate training data using POS ground truth. This innovation increased data throughput by 10x and slashed labeling costs, solving a critical scalability hurdle.
- Novel Out-of-Distribution (OOD) System: Invented a proprietary OOD detection system to identify "unknown" or non-inventory items. This breakthrough virtually eliminated false positives and misclassifications—a feature no other competitor could offer at the time.
- Massive-Scale Recognition: Architected a robust vision pipeline capable of distinguishing 10,000+ unique SKUs (designed to scale to 50k) with high precision in dynamic retail environments.
Operational Excellence & Edge Engineering
- Edge-Native Optimization: Engineered complex Transformer and CNN ensembles to run on constrained edge devices, slashing prediction latency from 5s to <300ms per transaction.
- Reliability at Scale: Built robust production pipelines and decentralized deployment workflows for remote updates across hundreds of locations, ensuring 99.99% system uptime.
Founding engineer for the ShopAssist Pulse analytics platform. Built the core perception stack from 0 to 1, enabling real-time customer journey tracking and store operational insights.
Retail Analytics & ShopAssist Pulse
- Privacy-Preserving Multi-Camera Tracking: Developed a novel tracking system to map customer journeys across overlapping and non-overlapping cameras without storing biometric data, ensuring full GDPR/CCPA compliance while delivering granular behavioral insights.
- Automated 3D Camera Calibration: Invented an online calibration method using environmental cues (like floor tiles and shelves) to auto-localize cameras in 3D space. This eliminated the need for manual checkerboard calibration and enabled accurate 3D localization from standard surveillance feeds.
- Efficient Inference Scaling: Led the migration to optimized inference engines (TensorRT), significantly boosting stream density per GPU. This optimization allowed for cost-effective scaling on commodity hardware, a critical factor for mass adoption.
Healthcare & Pandemic Response
- Multimodal Screening Kiosk: Rapidly architected and deployed a contactless screening solution combining Thermal Imaging, RGB Vision, and Gesture Recognition.
- Clinical-Grade Accuracy: Achieved medical-grade temperature detection for multiple subjects simultaneously. The system was validated by health professionals and played a vital role in protecting frontline staff during the COVID-19 crisis.
- Touchless Interface: Built a real-time hand gesture recognition system for touchless questionnaire answering, ensuring strictly verified safety protocols without physical contact.
Graduate Student Researcher
Tempe, Arizona • Jun 2018 - Sept 2019
Principal investigator for "DeepCrashTest", an NSF-sponsored project researching data-driven safety validation for autonomous vehicles.
- Adversarial Scenario Generation: Developed a generative framework using deep conv-nets to reconstruct pre-crash trajectories from dashcam videos, creating a library of "edge case" scenarios for AV testing.
- Simulation-to-Reality Gap: Bridged the gap between perception and control by feeding extracted trajectories into Webots, enabling repeatable and safe stress-testing of collision avoidance algorithms.
- Impact: Published at IEEE ICRA 2020. This work is cited as a foundational method for data-driven AV safety testing.
Key Projects & Systems
Building and deploying mission-critical AI systems at enterprise scale.
ShopAssist™: Frictionless Checkout
Architected an AI-powered checkout system capable of recognizing 10,000+ SKUs in real-time. Features proprietary Out-of-Distribution (OOD) detection to handle unknown items and uses vision-based barcode reading for >95% inventory coverage. Deployed at scale with <300ms prediction latency.
ShopAssist Pulse: Retail Analytics
Built the '0 to 1' perception stack for a real-time store analytics platform. Invented novel privacy-preserving multi-camera tracking and auto-calibration algorithms to map customer journeys and dwell times without expensive infrastructure changes.
DeepCrashTest: AV Safety Simulation
Developed a generative simulation framework that reconstructs adversarial crash scenarios from dashcam videos. Used deep convolutional networks to create data-driven 'edge cases' for rigorously testing autonomous vehicle collision avoidance systems.
Multimodal Healthcare Screening
Deployed a non-contact screening solution during COVID-19 combining Thermal Imaging and RGB gesture recognition. The system provided clinical-grade temperature checks and questionnaire answers, protecting frontline staff in major hospitals.
Research Publications
Pioneering work in Autonomous Systems and Generative Simulation, published in top-tier conferences.
DeepCrashTest: Turning Dashcam Videos into Virtual Crash Tests for Automated Driving Systems
S. K. Bashetty, H. B. Amor, and G. Fainekos
introduced a novel Generative AI framework that automatically parses dashcam footage to reconstruct pre-crash trajectories of multiple agents. This data-driven approach creates a diverse library of adversarial "edge cases" in simulation (Webots), enabling rigorous, repeatable safety validation of autonomous vehicle collision avoidance systems that was previously impossible with standard testing.
Patents & Innovations
Intellectual property driving the future of retail automation and computer vision.
Automatic item identification during assisted checkout based on visual features
RadiusAI, Inc. • Issued Apr 8, 2025
Core algorithms for extracting robust visual feature vectors from items at checkout. Enables precise identification of retail SKUs even under occlusion or varying lighting.
Point of sale station for assisted checkout system
RadiusAI, Inc. • Issued Feb 25, 2025
Hardware and software architecture for a next-generation POS station that seamlessly integrates computer vision to speed up checkout while maintaining human-in-the-loop verification.
Augmented reality of item identification during assisted checkout
RadiusAI, Inc. • Published Apr 10, 2025
Innovative AR system overlaying identification confidence and item details directly onto the cashier's view, reducing friction and training time.
Integration of visual analytics and automatic item recognition
RadiusAI, Inc. • Published Jul 25, 2024
Holistic framework combining item recognition with broader store analytics, linking checkout data to customer journey metrics for complete store intelligence.
Automatic item recognition from captured images
RadiusAI, Inc. • Published Jul 18, 2024
Real-time recognition methods optimized for edge computing, capable of handling rapid movement and diverse product orientations.
Visual analytics systems and methods
RadiusAI, Inc. • Published Jul 18, 2024
Broad analytics engine using video data to determine poses, gestures, and tracks, applicable to retail efficiency and healthcare screening.
Automatic item identification (Continual Learning)
RadiusAI, Inc. • Published Jul 24, 2025
Advanced learning pipelines for extracting embeddings, facilitating rapid retraining for new SKUs without catastrophic forgetting.
Automatic item identification systems
RadiusAI, Inc. • Published Apr 10, 2025
End-to-end system architecture for deploying scalable item identification across distributed retail networks.