Designed and built Dolcera's AI-native IP services marketing site end-to-end — visual design, copy, motion, and information architecture. Also shipped two print-ready 33″×81″ conference standees by reusing the site's React components and rendering them to vector PDFs through a Playwright-driven export pipeline. Stack includes a WebGL globe, custom SVG architecture diagrams, accessible Radix primitives, and a Zod-validated serverless contact endpoint.

An intelligent photo gallery system that curates daily photo selections using color harmony and semantic similarity. Built with CIELAB color space analysis and CLIP embeddings to select visually cohesive images, featuring an interactive gallery with modal inspection, color palette extraction, and dynamic background gradients. (try it below)

Distilled BirdNET into a 60.3 KB INT8 CNN running fully offline on an Arduino Nano 33 BLE Sense (Cortex-M4F). Benchmarked four model variants — scratch, transfer (MobileNetV2), knowledge distillation (T=4 soft targets), and INT8 post-training quantization — on 10 Pacific Northwest species. Deployed with live LCD readout and a Web Bluetooth BLE dashboard.

Benchmarked LSTM+Bahdanau, LSTM+Self-Attention, and Transformer architectures for multivariate time-series forecasting on ETTh1, Jena Climate, and Exchange Rate datasets. Evaluated trade-offs between attention complexity and prediction accuracy.

Data-driven fault detection and diagnosis on the Tennessee Eastman Process. Compared PCA, autoencoders, and optimized classifiers on 21 simulated fault scenarios across 52 process variables, achieving robust detection with interpretable diagnostics.

Tiny CNN Accelerator implemented on FPGA. Designed Verilog MFCC + 1D-CNN accelerator with AXI-Stream; timing closure at 100 MHz, <100 ms latency, real-time 16kHz audio. Achieved ≥90% accuracy on small-vocab Speech Commands with HPS integration, quantized weights, and on-board UI.