Dec 2024 – Present · 1 yr 3 mos · Sunnyvale, CA · On-site

• Amazon Lab126 Consumer Robotics
• Led and developed an end-to-end multimodal ML workflow enabling Amazon Astro Robot to interpret and respond to real-world queries using both visual and semantic data. Integrated on-device visual perception model knowledge and optimized prompts for MMLLM inference, enabling state-aware and person recognition question answering. Developed Android workflows (motion planning, camera, image upload) and a cloud-side LLM invocation with token streaming, slashing response latency from 15s to 4s and improving object state accuracy by 40%.
• Architected a pet monitoring workflow integrating Visual Perception and Presence API signals, cutting customer-facing false positive pet detections by ~90% and enabling a scalable, default-on feature rollout.
• Designed and implemented an Android AIDL-based SDK to enable interprocess communication of knowledge injection across robot subsystems. Standardized communication between device apps, and the centralized robotic memory graph, enabling 3+ Android applications to persist multimodal (vision + semantic) context at runtime, unlocked scalable ML real-time personalization workflows with robotics intelligence on embedded devices.
• Fine-tuned large action planning prompts for the robot across Claude Sonnet 3.5, 4.0, and 4.5 to align robot behavior with 20+ Amazon privacy policies for sensitive intent generation and data storage. Validated adherence through automated scenario scripts testing and manual testing, ensuring policy-compliant deployment of robot action planning in real-world home environments.

Feb 2022 – Dec 2024 · 2 yrs 10 mos · Sunnyvale, CA · On-site

• Spearheaded core development of Astro’s Home Monitoring platform, a flagship feature set enabling 24/7 sound, person, and pet detection, live view, and autonomous patrols—powering real-time security for more 10K+ households across consumer and small-medium business customers.
• Refactored Home Monitoring internal metric recording architecture and led org-wide migration to visualize metrics in AWS CloudWatch, delivered 30+ dashboards for all Home Monitoring features, standardized metric recording workflows, and accelerated onboarding through hosting cross-team training and documentation.
• Drove autonomous patrol success rate from 90% to 97% by performing time-series data analysis on S3 metrics from various levels like SLAM, mobility and floorplan to root-cause failures and lead a cross-functional fix initiative.
• Designed and implemented a time-based recovery mechanism in the Android monitoring application, reducing localization and system-related patrol failures by >30% and increasing reliability for 3,800+ daily active devices.
• Led design and implementation of partial patrol-point generation, reducing regeneration latency from 300s → 60s and improving patrol coverage stability for large (5,000+ sq. ft.) floorplans. Improved legacy point-placement algorithms in the Android stack with lifecycle-safe background execution and applied computational geometry techniques to maintain existing point coordinates, eliminating the long-standing issue where any floorplan edit forced a full point generation.
• Collaborated with PM and UX to ship a patrol timeline improvement, implementing an end-to-end error notification pipeline that enabled autonomous user troubleshooting and reduced related customer support tickets by 30%.