Wearable AI Memory for Meta Ray-Ban Glasses

• Built a real-time speaker identification pipeline for Meta Ray-Ban audio streams using Resemblyzer voice embeddings over a buffered WebSocket connection, dynamically mapping anonymous transcription labels to named contacts
• Designed a conversational knowledge graph using Apache AGE (PostgreSQL), utilizing LLMs for Subject-Predicate-Object (SPO) triple extraction to surface context-aware contact briefings via graph traversal

System architecture — glasses, iOS app, and backend processing pipeline

iOS app — contact knowledge graph and AI-generated recap

CUDA Attention Kernel

• Implemented a tiled CUDA forward attention kernel featuring online softmax and shared-memory blocking, shifting arithmetic intensity from 7.5 to 131 FLOPs/byte to successfully transition the workload from memory-bound to compute-bound
• Validated numerical correctness against PyTorch scaled_dot_product_attention across causal and non-causal paths within FP32 precision limits (zero absolute error); verified 99.8% of GPU time spent in-kernel via profiling

colab.research.google.com

MiniAttention — Benchmark Notebook

Colab notebook profiling the custom CUDA attention kernel against PyTorch's scaled_dot_product_attention.

Multilingual Children's Book Studio

• Shipped a concurrent asset engine pairing a Go API with Celery/Redis workers to orchestrate parallel text, image, and voice generation
• Implemented hybrid search pairing Elasticsearch and vector databases for semantic content recommendations in production

Worker pipeline — Redis-queued Go workers for indexing, translation, and audio

Audio worker detail — Google Cloud Speech API, Cloudinary, and MongoDB

AI Patent Analytics Platform

• Led a 6-developer team to build a full-stack MERN application for automatic document parsing and patent analysis using GPT-4 with custom prompts, with a data pipeline to process 108GB of USPTO public patent application data for LLM training
• Architected microservices for parsing PDFs to JSON, feeding custom prompts into LLM for inference, and generating PDF reports

IPELiNT — Intelligent IP Analytics Done Efficiently

TCU SciCom research poster — architecture, LLM workflow, and results