Deep dives on flagship work.

FIRST Agentic CSA is an MCP server that gives FRC teams a single, trustworthy place to search across the entire documentation ecosystem. Instead of hopping between WPILib, REV, CTRE, Redux Robotics, and PhotonVision, students can ask plain English questions and get documentation-grounded answers that respect the authoritative sources teams rely on during build season.

The project is designed to integrate directly with VS Code and AI coding assistants so that documentation lookups happen where teams already work. It supports filtering by programming language and documentation year, and ships with a ready-to-drop copilot instructions file that nudges assistants to search docs before answering. The result is faster onboarding, fewer wrong answers, and a shared resource for 3,600+ FRC teams under intense time pressure.

FRC Shooter Selector is a physics-first simulation toolkit for the 2026 REBUILT game that helps teams design shooter mechanisms before building expensive prototypes. It models full ballistics with quadratic drag on foam game pieces, then translates the physics into practical motor, gearing, and flywheel recommendations teams can implement quickly.

The tool provides conservative torque calculations that avoid overheating, sensitivity analysis for angle and velocity errors, spin-up time comparisons across gear ratios, and flywheel mass tradeoff studies. Teams can use the browser-based app, a Django web interface, or the CLI for batch analysis. Outputs include Plotly charts, ready-to-copy RPM lookup tables, and PNG plots for design reviews.

Rubik TFLite is a hands-on demo that calls TensorFlow Lite using the libQnn delegate as an external delegate, defaulting to the HTP backend for QNN. I worked with a student to package the workflow so teams and developers can run on-device image transformation models without the usual setup friction.

The project ships with C and Python demos that load input images from disk, run TFLite inference through QNN, and write outputs back to the filesystem. It targets image transformation models and highlights how to wire TFLite to Qualcomm hardware in a reproducible, student-friendly flow.

Example runs use the Real-ESRGAN x4+ model from Qualcomm AI Hub and show end-to-end invocation with sample inputs and outputs. Supported input formats include JPEG, PNG, TGA, BMP, PSD, GIF, HDR, PIC, and PNM, with PNG output for easy inspection.

Deployed overseas with a HIMARS maintenance company and served as a local interpreter across multiple Middle East deployments. Started as a 91B mechanic, then moved into logistics and platoon leadership.

Led a 35-soldier maintenance platoon, automated tracking to remove bottlenecks, and earned the Army Commendation Medal for complex repairs and readiness gains. Known for mastering technical manuals, diagnosing hard failures fast, and coaching teams under pressure.

Partnered with the Designing for Defense program on Navy equipment tracking, consolidating data from multiple systems into commander-ready reports.