Accelerated Ternary Speech Intent Recognition on FPGA
April 2026
https://github.com/bencejdanko/speech-to-intent-cognichip
Mobilint MLA100 NPU Computer Vision Applications
March 2026
https://github.com/bencejdanko/imagenet-classification-mobilint-mla-100
  • Worked with Mobilint hardware and compiler engineers to evaluate computer vision applications for the MLA100 NPU.
  • Developed and optimized a real-time image restoration pipeline for the NPU, achieving <6ms latencies for 256x256 images.
  • Replaced a high-latency ResNet baseline (400ms+) with a custom 0.2M parameter Convolutional Network by optimizing kernel launches and restricting feature maps to 64 to minimize hardware-induced latency.
  • Orchestrated large-scale training and data processing jobs using modal, scaling cloud GPU resources to conduct progressive training cycles with Cosine Annealing and mixed precision.
  • Optimized PSNR improvement through iterative checkpoint-based training and rigorous hardware-aware model profiling.
  • Evaluated and compiled ImageNet classification models for the MLA100, comparing inference throughput against Intel AI Boost and achieving 2.7x performance gains.
Edge Vision Smart Elevators: Privacy-Preserving Headcount Analysis
March 2026
https://github.com/bencejdanko/m55m1-ElevatorCounting-YOLOv8n
  • Conducting applied research for Nuvoton’s NuMicro M55M1 microcontroller, architecting a confined person-recognition pipeline achieving 15 FPS within 1.5MB SRAM constraints.
  • Engineered an INT8-quantized YOLOv8n pipeline using the Vera compiler and Arm Ethos-U55 NPU.
  • Curated a specialized dataset of 13,000+ images from Roboflow, and distributed for common public use on Huggingface.
  • Assembled a proprietary data collection setup using a Raspberry Pi Zero W and 160° FOV camera for real-world confined elevator environments.
  • Developed custom firmware using CMSIS-CSolution to satisfy business presentation requirements.
  • Collaborated with Nuvoton hardware engineers to validate system performance for a live demonstration at the Sensors Converge conference.
Full Stack RBAC Portal and Data Migration for Inclusive World
November 2025
https://github.com/bencejdanko/inclusive_world_portal
  • Led a team of 8 undergraduate developers as the Full Stack Lead for Inclusive World, a non-profit dedicated to empowering people with disabilities.
  • Architected and implemented a robust Role-Based Access Control (RBAC) management portal for directors to streamline volunteer and member coordination.
  • Orchestrated a critical migration from legacy Google Sheets-based automations to a professional-grade tech stack, iterating in designs from React/Supabase to Django and PostgreSQL.
  • Engineered comprehensive data models to handle complex member relationships, replacing manual spreadsheet workflows with automated Python-based backend logic.
  • Managed the end-to-end development lifecycle over 8 months, ensuring code quality, team synchronization, and goal alignment with organization directors.