At ShelfEx, I built distributed crawlers for ShelfIntel that process 10M+ records a day and expanded coverage to 27M+ products. I also built headless collectors and recovery paths for changing APIs and anti-bot challenges.
For ShelfIntel, I built GCP data pipelines that cut infrastructure cost by 95%, along with an XGBoost model to estimate product sales velocity from limited observable signals.
I built Chanakya, a stateful orchestrator–worker runtime coordinating six worker agents across long-horizon AI workflows. Its persisted task state, context compaction, and resumable execution support continuity across agents.
For ShelfPulse, I built a VLM/OCR pipeline that achieved 95–99% structured extraction accuracy across 2,000+ menus. I also expanded labeled data from 500 to 50,000 images and deployed Vertex AI inference processing 50,000+ images per week. On Solara AI, I built a video-processing system that turns screen recordings into product-demo videos using FFmpeg and OpenCV.
