Projects

End-to-end pipeline for fine-tuning a small LLM on structured JSON extraction tasks, then deploying the result on-device. Used LoRA adapters on Llama 3.2 3B via MLX (Apple Silicon) and evaluated across four artifacts: generic F16, generic Q4_K_M, fine-tuned F16, and fine-tuned Q4_K_M. The deployable artifact — fine-tuned Q4_K_M GGUF — achieved 0.961 recall and 100% JSON validity at 2.91s average latency on llama.cpp, compared to 0.780 recall and 6.65s for the generic F16 baseline. Quantization after fine-tuning cost only 0.011 recall while cutting latency in half. Evaluation ran on a 106-case held-out set with automated recall, JSON validity, and schema validity scoring.

The Halo is a mandatory titanium arch on all F1 cars. It saves lives but cuts through the most interesting part of onboard footage. This project removes it cleanly from visor-cam video using two stages. Stage one: classical CV mask detection — Sobel-Y gradient detection for the arch edge, a robust probe-and-fit keel detector with outlier rejection and temporal jump guards, and explicit geometry construction to avoid over-masking. Stage two: two inpainting methods compared side by side — LaMa (Fast Fourier Convolution network, per-frame spatial inpainting) and RAFT optical flow with backward warp and distance-transform blending for temporal coherence across 300 frames at 60fps.

Built a complete computer system from the ground up as part of the Nand2Tetris course (ECEG 431). Starting with a single NAND gate, I designed and simulated all hardware components including logic gates, ALU, RAM, and the CPU. On the software side, I developed an assembler, a virtual machine translator, and a compiler for a high-level object-oriented language, culminating in a fully functional Operating System. This project provided a deep, demystified understanding of how computers actually work under the hood.

Aggregates coffee shop data from various sources and displays it in a user-friendly interface. Use it to find your next brew!

Uses OCR and PDF parsing to extract text and convert it to Markdown, for easy editing and formatting.

Architected full-stack web application for managing HRI experiments with real-time robot control interfaces. Implemented WebSocket-based bidirectional communication protocols for low-latency robot teleoperation. Designed RESTful API leveraging Robot Operating System with JSON-defined plugins for extensibility across multiple robot platforms. Created comprehensive logging system capturing interaction data, timestamps, and experimental conditions for reproducibility. Technologies: Next.js, React, TypeScript, Node.js, WebSockets, PostgreSQL, Docker.

Designed embedded control system for autonomous hydrogen fuel cell-powered vehicle using finite state machine architecture. Implemented real-time sensor fusion combining spectrometer readings and power monitoring with calculated stopping algorithms. Developed PlatformIO-based build system with hardware abstraction layer for testing and simulation. Achieved precise distance control (±10cm) through chemical reaction timing at AIChE National Competition. Technologies: C++, Arduino, PlatformIO, I2C/SPI protocols, finite state machines.

Developed ensemble machine learning models (LightGBM, XGBoost, Random Forest) to predict F1 lap times with high accuracy. Engineered features from weather data, track characteristics, and historical performance using domain knowledge. Implemented cross-validation and hyperparameter optimization for model evaluation across multiple racing circuits. Analyzed feature importance to understand factors influencing racing performance. Technologies: Python, LightGBM, XGBoost, Random Forest, pandas, scikit-learn, FastF1 API.

Built production system serving 1500+ concurrent users and 250k+ monthly visitors. Implemented WebSocket-based real-time data streaming with automatic reconnection and state synchronization. Designed responsive UI with accessibility features meeting WCAG 2.1 AA standards. Optimized database queries reducing page load times by 60% through intelligent caching and indexing. Technologies: Next.js, TypeScript, PostgreSQL, Docker, DigitalOcean.

Developed a full-stack invoicing platform prioritizing reliability, security, and professional user experience. Features include multi-step invoice creation with flexible line items, automated PDF generation, client management with complete contact details, business profile management with branding support, and comprehensive status tracking (draft → sent → paid/overdue). The application implements proper authentication, input validation, and error handling throughout, with a mobile-first responsive design using shadcn/ui components.

Designed and developed a personal portfolio website using modern web technologies. Features include responsive design, dark mode support, PDF rendering for CV display, and a clean, professional interface for showcasing projects and experience.

This page details the importance of web accessibility and the specific features implemented on this website to ensure an inclusive experience for all users, including those with disabilities. It documents the accessibility practices, standards compliance, and testing methodologies used.

Created an accessible tutorial video explaining LaTeX, a typesetting system commonly used for technical and scientific documents in academia. The video covers how to set up Overleaf as an editor, explains key LaTeX tags and formatting, demonstrates both inline and display math equations, and provides a complete walkthrough of creating your first document with proper formatting.