Three Tracks, One GPU: Lessons from the AMD DevMaster Hackathon

What happens when developers are given the freedom to build across three rapidly evolving areas of AI? At the AMD DevMaster Hackathon, nearly 3,900 participants took on the challenge across Multimodal AI, Agentic AI, and Physical AI, using AMD Radeon™ GPUs and the AMD ROCm™ software to turn their ideas into working applications. Looking across the submissions, three clear lessons emerged.

Lesson 1: The best multimodal applications solve a specific problem

The Multimodal AI track challenged developers to combine text, images, video, audio, and visual generation into practical experiences. With over 1,500 registrations, it was the second-largest track. The strongest submissions focused on applying multimodal capabilities to clearly defined user needs. The top three winning projects included:

• 1^st Place: Team Rivet developed an intelligent advertising tool that helps small and micro businesses create and review commercial advertisements.
• 2^nd Place: Team austin1997 created an accessibility tool that automatically generates voice narration for videos, helping make video content more accessible to visually impaired users.
• 3^rd Place:  Team N DIVIJ built a voice-driven storybook creation tool that turns spoken input into complete children’s story videos.

The other five teams among the top eight received Excellent awards: team Junjun Liu, team icuic, team peace_&_love, team panbu2007, and team gaojie.

The lesson: Multimodal AI becomes more valuable when multiple capabilities come together to solve a clearly defined problem. The submissions also demonstrated that Radeon GPUs can provide capable local compute for these workloads, helping developers build responsive experiences while keeping processing closer to the user.

Lesson 2: Effective agents need to reason, act, and verify

The Agentic AI track attracted the largest field, with over 1,700 approved registrations. Participants built local agents capable of reasoning, tool use, memory management, and task execution, with a strong emphasis on Radeon GPU and ROCm optimization. Top submissions treated inference and execution as core design constraints. Examples included a multi-agent financial research system on a ROCm-optimized runtime, a coding agent built with a custom HIP inference engine, and a QA agent combining retrieval, rule-based logic, automated testing, and human validation.

• 1^st Place: Team SignalForge Labs built a locally deployed financial research workbench that gives individual investors a private environment for AI-powered research and analysis.
• 2^nd Place: Team Aetheris Blackbox showcased a fully localized programming agent that supports both desktop and terminal environments.
• 3^rd Place: Team Xieweikai created an enterprise customer service compliance agent that uses AI to inspect interactions and identify potential compliance issues.

The other five teams among the top eight received Excellent awards: team DarthCeltic, team 2SIN, team abidedavana, team Hipscope, and team himanshu748.

The lesson: Effective agents require more than reasoning ability. They must also execute reliably, run efficiently on local hardware, and produce verifiable outputs.

Lesson 3: Physical AI connects intelligence to the real world

The Physical AI track had over 500 registrations, spanning robotics, simulation, navigation, and embodied intelligence. A key theme among the strongest projects was sim-to-real: connecting what an AI system learns in simulation with how it performs in the physical world. Several projects also contributed to upstream open-source robotics ecosystems, reinforcing the importance of shared infrastructure.

• 1^st Place: Team Binh Pham demonstrated a full-stack physical simulation, rendering, and reinforcement-learning environment running on a single Radeon GPU, bringing training and inference together for physical AI workloads.
• 2^nd Place: Team KINESYS developed a ROCm-native humanoid robot manipulation pipeline and demonstrated it through a real-world task involving retrieving tableware from a dishwasher.
• 3^rd Place: Team Wenjie Ouyang introduced a solution that generates robot motion-control policies from monocular human videos, translating observed human movement into robotic behavior.

The other five teams among the top eight received Excellent awards: team Enoch, team Akbro23, team Yuhao Cao, team Robotics Gemini, and team Phi Media Lab.

The lesson: Physical AI requires more than an intelligent model. Simulation, compute, robotics software, and real-world behavior must work together.

Three tracks, one common theme

Across all three tracks, one theme was consistent: practical AI is no longer just about model capability. Developers are building systems that create, reason, act, and interact with the world, while treating hardware and software as part of the application itself.

Ready to keep building?

Join the AMD AI Developer Program to access cloud credits, free premium training, and a community of developers building the future of AI on AMD hardware.

The hackathon may be over, but the work it showcased is only beginning.