Designed. Built.
Validated.
Designed and programmed an autonomous differential-drive robot using ROS2 Humble to complete navigation, cargo handling, and rescue tasks for the Peter Gregson Design Challenge.
- Integrated a Beelink MiniPC running Ubuntu 22.04 with a custom ATmega328P-based PCB for precise hardware control.
- Implemented a distributed ROS2 architecture using /cmd_vel, /scan, and /odom topics for SLAM and costmap-based path planning.
- Engineered a custom LED safety system to broadcast spatial threat levels dynamically based on 360° LiDAR data.
Real-world object detection isn't a benchmark — it's rain, mud, motion blur, and uneven lighting on machinery moving through a live field. This system had to work in those conditions.
- Benchmarked and fine-tuned YOLO, D-Fine, and RFDETR — optimizing each model's speed/accuracy trade-off on edge hardware
- Handled real field challenges: occlusion, variable natural lighting, non-uniform crop shapes, and continuous motion artifacts
- Built data pipelines, logging infrastructure, and evaluation frameworks to track model performance across conditions
An award-winning trailer maneuvering and storage system — designed to support loads up to 3,175 kg while remaining user-operable without heavy machinery. Every decision was backed by analysis, not instinct.
- Full structural analysis at SF=2, computing member stresses and validating geometry through method of sections
- Complete 2D/3D CAD in SolidWorks including assemblies, Bill of Materials, and real supplier cost analysis
- Prototyped and physically tested — evaluating stability, alignment, and load distribution under simulated conditions
Designed analytically and validated physically — not just simulated. The gap between theoretical and actual performance told a better story than the numbers alone.
- Applied method of joints and sections to calculate all member forces and safety factors before fabrication
- Validated through progressive physical load testing — bridge held 368% beyond its calculated capacity
- Demonstrated the value of conservative design margins and precise fabrication in structural engineering
A Qt/C++ desktop application that streams continuous numerical data over TCP/IP and renders it live in a scrolling graph. The challenge wasn't the UI — it was making TCP behave like a reliable message protocol when it fundamentally isn't.
- Implemented a size-prefixed framing protocol to safely reconstruct data samples even when TCP packets arrive fragmented or batched
- Transmitter (QTcpServer + QTimer) and receiver (QTcpSocket + QCustomPlot) run as one demo window or two separate processes simulating real networked machines
- Multi-client support, configurable rolling time window with auto-scaling Y-axis, and optional CSV logging of timestamped samples
Built out of necessity — after spending too many hours in labeling tools not designed for how I actually work. A full desktop annotation application in PyQt5 for preparing object detection training datasets at scale.
- Custom canvas widget with click-and-drag bounding box drawing, coordinate scaling between display and original image space, per-class color coding
- Three export formats: YOLO .txt (normalized), Pascal VOC .xml, and cropped image patches at 320×320 or 640×640 — ready to feed directly into a training pipeline
- Batch folder loading, auto-save to JSON on every image switch (no annotation loss on crash), and a full class manager with persistent color assignments
Trivago for food delivery. The real cost of an order isn't the menu price — it's menu + delivery + service fees + surge. FoodFinder Pro compares all-in totals across DoorDash, Uber Eats, and SkipTheDishes, then deep-links you to the cheapest checkout.
- Core screens in React Native (TypeScript, Expo): Search, Price Comparison, and Pro Account — with typed quote models and platform metadata normalization
- Custom hook (useDeliveryQuotes) with short-TTL in-memory caching; prices stored in minor units (cents) to avoid floating-point currency errors
- Price alert system with local persistence via AsyncStorage — users set a target total and track when it drops
