Squeezing Maximum AI into Minimal Hardware

Highlights from the Minimalist Edge AI Challenge By IEEE Industry & Computing Chapter

Deploying advanced computer vision models to resource-constrained edge hardware is one of the toughest engineering hurdles facing embedded systems today. At the SEC 2026 Minimalist Edge AI Challenge hosted at Santa Clara University, student innovators demonstrated how to bridge this gap. 17 teams from 7 universities tackled practical edge deployments on compact boards like the Arduino UNO Q and Qualcomm development hardware, with participants proving that strict memory and power ceilings do not have to compromise real-time performance.

Hardware Track

For the Hardware Track, student teams were challenged to design, optimize, and deploy complete, end-to-end edge AI applications directly on physical Qualcomm Arduino UNO Q hardware. Evaluated on problem innovation, system architecture, on-device latency, RAM footprint, energy efficiency, and live demonstration performance, these projects demonstrated practical breakthroughs in real-world, constrained hardware deployment

The Big Takeaway: Layered Optimization Beats Isolated Fixes

Running real-time object detection locally requires a cohesive optimization stack rather than a single adjustment. First-place winners of the hardware track, Edge Visionaries, demonstrated that successfully running YOLO11n on edge devices hinges on simultaneously combining resolution scaling, INT8 quantization, and CPU multithreading. By restructuring the data execution pipeline to alleviate memory bottlenecks, teams can deploy low-latency vision models for restricted industrial safety zones and hazardous work environments without relying on the cloud.

Team: Md Nafis Tahmid Akhand (Louisiana Tech University), Md Mahfuzul Haque Gazi (Louisiana Tech University), Komala Subramanyam Cherukuri (University of North Texas) (shown here alongside judges and contributors to the event.)

Runners-up Provided an Important Focus on User Needs

Second Place: Tongtong Wang (Santa Clara University) and Victor Joulin-Batejat (Santa Clara University) created Luxo III, an emotive robotic arm capable of real-time interactive responses based on user gesture recognition.

Third Place: Samrithi Balaa Saravanan (Santa Clara University), Nidia Jupudi (Santa Clara University), Faith Ogunfunmi (Santa Clara University), and Sunjana Banwait (Santa Clara University) created SafeCue, a low-power, on-device environmental sound classifier that detects safety-critical sounds (smoke alarms, doorbells, sirens) and triggers an instant visual alert for deaf and hard-of-hearing users.

Software Track

For the Software Track, student teams were challenged to optimize a heavy baseline model using a provided dataset, aiming to maximize efficiency while minimizing accuracy loss. Evaluated on accuracy, computational throughput, and energy efficiency across real-world edge scenarios, these projects demonstrated practical breakthroughs in lightweight machine intelligence.

The Big Takeaway: Aggressive Compression Beats Surface Tuning

Optimizing heavy AI workloads for resource-constrained edge environments demands an aggressive compression strategy rather than surface-level tuning. First-place winners of the software track, Quantax, proved that shrinking complex models without compromising utility relies on combining targeted structural pruning with post-training INT8 quantization. By eliminating redundant neural parameters and streamlining execution graphs, software-level optimizations can slash memory footprints and inference latency across real-world edge tasks, enabling ultra-lean intelligence directly on minimalist microprocessors without requiring dedicated accelerator hardware.

Team: Yefeng Yuan (Santa Clara University), Yulin Zeng (Santa Clara University), Scott Wang (Santa Clara University)

Runners-up Showcased End-to-End Optimization Pipelines

The second place winners Jacob Cho (University of Southern California), Jungoo Kim (University of Southern California) created Chilly Ships, a structured software optimization pipeline combining block removal, Taylor-importance channel pruning, distillation recovery, and static QDQ INT8 quantization to slash MobileNetV2 model size by 88% and latency by 80% while preserving a 97.39% validation accuracy.

Third place winners Weiyao Chang (Santa Clara University), Yuanzhen Yang (Chung Yuan Christian University) created Valiant Joy, a modular edge computer vision framework that leverages structured pruning, INT8 quantization, and ONNX Runtime execution cores to deliver low-latency benchmarking and inference across classification, detection, and tracking tasks.

Industry Judges and Support from IEEE CS Leadership

Qualcomm and STMicroelectronics sponsored hardware boards and provided engineers on site to mentor the hardware teams while NihaoPay provided awesome swag for the students. Engineers from all three companies also joined as judges. About 15 professionals from industry and academia contributed to the event by leading pre-competition webinars, developing judge rubrics, and serving as on-site judges.

The awards were presented to the winners by Avery Lu (IEEE Santa Clara Valley Section Chair) and Andreas Reinhardt (IEEE Computer Society VP for T&C Board).

Attendee Perspective

“Optimizing AI workloads for resource-constrained hardware has always fascinated me. My research focuses on designing compilers that automatically generate optimized execution pipelines for AI models on FPGAs under specific resource constraints. Working on this challenge was an exciting opportunity to explore these ideas from a practical perspective. I got to work with the UNO Q, a compact yet surprisingly capable board, and gained valuable hands-on experience navigating the hardware and resource constraints of edge AI systems. It was a rewarding experience to bridge my research interests with real-world hardware challenges.” — Md Mahfuzul Haque Gazi (Louisiana Tech University), Hardware Track, First place team member

"We went in thinking the hard part would be the compression. It turned out to be knowing which result was real. The model we submitted came from measuring the recovery, not the pruning — and that’s the lesson we’re taking with us." — Yuanzhen Yang (Chung Yuan Christian University), Software Track Finalist, Team Valiant Joy (Chung Yuan Christian University)

Share Your Story through IEEE Computer Society

Do you have your own success story to share from an event supported, sponsored, or run by IEEE Computer Society? Send it our way by emailing us at awesome@computer.org. Want to get involved? Check out our SYP Mentoring program or check out your local chapter.