Selected Projects
Here's a glimpse into some of the challenges I've tackled and solutions I've built.
Water Accounting IoT Platform
Low-power telemetry + ledger architecture enabling growers to comply with California SGMA in real time.
Virtual Power Plant Toolkit
Python library & dashboard for optimizing battery dispatch under NEM 3.0—cut residential bills by 30%.
Amazon Last-Mile ML Experiment
Led a pilot blending RL & heuristics that cut route planning compute cost by 15%. (NDA-friendly summary)
Water Accounting IoT Platform
This platform addresses the critical need for precise water usage tracking for agricultural producers under California's Sustainable Groundwater Management Act (SGMA). It combines low-power, long-range (LoRaWAN) telemetry devices in the field with a serverless backend on AWS. The system provides a verifiable ledger of water rights usage, offering growers peace of mind and robust data for regulatory reporting.
Key outcomes include reduced manual reporting overhead, improved data accuracy for water management decisions, and a scalable architecture ready for wider adoption.
← Back to ProjectsVirtual Power Plant Toolkit
With California's NEM 3.0 significantly altering solar export compensation, this Python-based toolkit helps homeowners with battery storage systems (like Tesla Powerwalls) optimize their energy usage. It ingests historical consumption data, solar generation patterns, and complex utility tariffs to model optimal battery dispatch strategies. The goal is to maximize self-consumption and minimize electricity bills under the new regulations.
Initial simulations and real-world tests have shown potential residential bill reductions of up to 30% by strategically charging and discharging the battery based on time-of-use rates and demand charges.
← Back to ProjectsAmazon Last-Mile ML Experiment
During my tenure at Amazon, I spearheaded a pilot project aimed at optimizing computational costs associated with last-mile delivery route planning. This initiative explored a hybrid approach, blending reinforcement learning (RL) techniques with established heuristics to find more efficient routing solutions without compromising on delivery times or accuracy.
The experiment successfully demonstrated a ~15% reduction in compute costs for the tested scenarios, offering a promising avenue for significant operational savings at scale. This summary is intentionally high-level to respect confidentiality agreements.
← Back to Projects