Projects
Selected work across full stack web apps, improving AI efficiency, embedded systems, machine learning, AI for education, and processor design.
Study tool
Study Tracker
Developed and deployed a responsive full-stack study journal using Next.js, TypeScript, Prisma, PostgreSQL, and Vercel. Implemented credential-based authentication, study-session CRUD, journal entries, a persistent focus timer, weekly reflections, and interactive calendar and Recharts analytics for reviewing study patterns across multiple timeframes. Added server-side caching, secure password hashing, and reliable per-user data persistence.
Embedded audio
Programmable Multi-Effects Guitar Pedal
Built a Daisy Seed-based programmable guitar pedal with selectable bypass, distortion, reverb, delay, and flanger effects. Programmed real-time embedded C++ DSP firmware using DaisySP, audio callbacks, delay-line processing, wet/dry mixing, and debounced mode switching. Soldered a perfboard prototype integrating guitar input buffering, audio jack wiring, and control inputs.
Embedded Software
Reusable Embedded C Library
Developed a reusable embedded C library with 5 core modules: GPIO, UART, ring buffer, FSM, and button debouncing using defensive input validation and static/caller-provided memory for embedded environments. Designed a platform abstraction layer separating hardware-independent GPIO/UART APIs from STM32G071 HAL implementations, translating portable driver types and status codes to platform-specific HAL operations. Implemented an O(1) circular byte buffer, configurable time-based button debouncing with press/release events, and a table-driven state machine that validates states, events, and duplicate transitions before execution.
Machine Learning, Data Processing
Guitar Audio Classification Model
Built an audio classification pipeline with 89% test accuracy to identify four guitar sound categories: single notes, chords, palm-muted playing, and background noise from self-recorded audio. Processed one-second WAV clips and extracted MFCCs, RMS energy, spectral centroid, bandwidth, rolloff, and zero-crossing-rate features. Trained and evaluated a logistic regression classifier using multi-session training, validation, and test splits on independently recorded guitar samples.
Computer Architecture / Machine Learning
CPU Optimized Inference Engine (WIP)
CPU neural-network inference engine in C++ from scratch. Built a CPU neural-network inference engine from scratch in C++, implementing matrix multiplication, dense layers, activations, convolution, and model execution. Optimized performance on Apple Silicon through cache-aware tiling, ARM NEON SIMD, multithreading, and INT8 quantization, using reproducible benchmarks and profiling to compare against Apple Accelerate.
Machine Learning
Post-Disaster Building Damage Classifier
Developed a CNN-based computer vision pipeline to classify structural damage from pre- and post-disaster satellite imagery across four severity levels, using image preprocessing, augmentation, and transfer learning. Built an evaluation pipeline using precision, recall, F1 score, confusion matrices, and class-distribution analysis to diagnose model performance on imbalanced disaster-damage categories. Implemented visual inference outputs that overlay predicted damage severity on satellite imagery to support rapid identification of heavily damaged areas following natural disasters.
Digital logic
RISC-V RV32I Single-Cycle Processor
A Verilog single-cycle processor with program counter, register file, ALU, immediate generator, instruction memory, and data memory. In addition, there are testbenches verifying behavior of each independent module. Entirely written in Verilog.
Image processing
Seam Carving Web App
Built and deployed a full-stack image resizing web app with upload previews, axis controls, projected dimensions, and downloadable results. Engineered an API route that validates PNG, JPEG, and WebP uploads up to 12 MB, creates isolated temporary jobs, invokes the Python processor, enforces timeouts, and cleans up files. Implemented content-aware resizing using Sobel energy maps, dynamic programming, and seam backtracking to remove low-energy pixels while preserving image content.
Backend / System design
URL Shortener
Built a URL shortening API with Python, Flask, PostgreSQL, and Redis. Generated compact Base62 short codes from database IDs and implemented HTTP redirects with Redis cache lookups, PostgreSQL fallback, and cache repopulation. Containerized the application, database, and cache with Docker Compose, including persistent database storage, service health checks, and a health endpoint.
Hackathon Projects
AI efficiency
EcoToken
A Chrome extension that uses a local prompt-complexity classifier to route each Gemini prompt to the lowest-cost model tier that can handle it, with an analytics dashboard showing estimated inference cost savings and CO2 emissions prevented.
Education
TeachBack: Inverted Tutoring
A hackathon tutoring app where students teach an AI student, with memory-aware personalization and step-gated problem solving.
Tools I Use
- C++
- Python
- TypeScript
- HTML/CSS
- Verilog
- Git
- Docker
- Vivado
- Next.js
- ESP32
- STM32
- STM32CubeMX
- Soldering
- PyTorch
- TensorFlow
- scikit-learn