Austin, TX

• •Developed STM32H7 FreeRTOS firmware for Toshiba medium voltage drives; built comprehensive testing infrastructure from zero with unit/integration/HIL tests (Python, Bash, Ruby). Reduced QA team workload by approx. 60% by creating production CI/CD pipeline validating 10,000+ parameters , running all unit, integration, and HIL tests in around 10 minutes.
• •Optimized RTOS task priorities and DMA scheduling, reducing TouchGFX CPU overhead from 41% to 18% , achieving 25% total system CPU reduction , and eliminating all timing violations to correct issues revealed by my new HIL tests. (C, C++) .
• •Developed TouchGFX UI screens with inter-task communication using CMSIS osMessageQueue for telemetry transfer between tasks. Double-buffered 30 FPS rendering with sub 80µs UI execution time ensures high-priority operations arent blocked.
• •Engineered a RAG based AI agent using Azure Copilot and OCR for detecting defects in CAD drawnings with 94% precision.

Austin, TX

• •Collaborated with research team funded by Texas State C.A.D.S to fine-tune Vision Transformer and MASK R-CNN models on Red Hat Linux SLURM cluster with Nvidia A100 GPUs. Created custom dataset achieving 98% precision for defect detection
• •Built production REST API using Python FastAPI, PostgreSQL, and Docker for deployment on Huggingface; implemented server side async request handling and batch processing to handle concurrent requests from React Native mobile app
• •Accelerated inference via layer fusion and ONNX to TensorRT engine conversion; reducing latency and cloud costs by 40%

San Marcos, TX

• •Developed cross-modal distillation pipeline (video to IMU sensors ) and personally trained custom multimodal transformers for fall detection under Dr. Anne Ngu. Deployed to edge devices ( phones , wearables ) with 92% F1 score in real time
• •Built custom CUDA preprocessing kernels in C++ (windowing, normalization, FIR) w/ 11x speedup vs numpy, parallelized training via PyTorch DDP, tested DSP algorithms to clean dataset and reduce noise, improving F1 score by 10%
• •Refactored multimodal PyTorch / TensorFlow transformer models to edge via TFlite and ONNX using INT8 quantization + mixed-precision training, achieving 2-3x battery life w/ sub-1% accuracy loss allowing for on device inference ( Python ) .
• •Refactored full-stack Android app (Java / Kotlin) to support ONNX , AWS S3 , and DynamoDB for sensor or user analytics. Shipped features like dynamic input support / pre-processing configs, async, to reduce noise and improve reliability