Projects

Building an AI-native platform that transforms how small and medium businesses make decisions. The system connects multimodal business data — chat logs, emails, documents, transaction records — into a semantic data layer that models real-world entities and relationships, encodes domain-specific business logic from veteran operators, and enables autonomous AI agents to execute routine tasks and generate strategic recommendations.

Selected as a winner at the SDx & Cadre AI OpenClaw Showcase in Southern California, presenting as "Personal Palantir".

Commissioned by a manufacturing export company to rebuild their digital operations from the ground up. Spent the first two weeks embedded on-site — interviewing the CEO, department managers, and assembly-line workers — to map every pain point: fragmented document management, manual quoting workflows, disconnected client tracking.

Applied an agile framework with weekly sprint cycles over five months to architect and deliver a production CRM/ERP system. Built end-to-end order flows for customer management, quotes, invoices, purchase orders, and payment tracking with PDF generation and email integration.

Led a 7-person engineering team to build a production system that processes terabytes of messy GPS data to compute real-time transportation carbon footprints. Deployed AI models via FastAPI and Docker, built a React Native mobile app with item-based recommendation engine, and enforced strict data privacy compliance for US and China users. Served 300+ active users.

Built an end-to-end AIoT checkout system integrating camera-equipped smart scales with a cloud-based inference backend. Developed a CNN classifier optimized with Bag-of-Visual-Words (BoVW) for dish recognition. Engineered real-time pipeline: edge image capture, cloud inference, and API callbacks for pricing in under 100ms.

Discovered a 14% increase in the day-to-day variability of the Southern Hemisphere's dominant climate mode, driven by thermodynamic shifts that fuel more extreme weather events. Published in Geophysical Research Letters.

Can we detect transportation modes using only speed data? Built a distributed ETL pipeline with Dask to process 4TB of raw GPS trajectories, then designed a privacy-preserving Transformer model achieving 96% SOTA accuracy. Demonstrated robust cross-regional generalization, boosting accuracy from 80.5% to 86.1% across datasets.

Designed a Federated Learning system for optimizing prompts across heterogeneous black-box LLM architectures, addressing data privacy and communication constraints in edge computing. Published as a system demo at ACM MobiCom 2024.