@josh05naim26281

Malatang is a measured self-improvement harness for bug-fixing agents. Instead of claiming that an AI “gets better” based on vibes, Malatang defines improvement operationally: run the same frozen benchmark, accept fixes only when build and tests pass in a sandbox, record pass rate, and let the system modify its own strategy between runs. The benchmark contains 25 seeded training bugs across syntax errors, off-by-one logic, null handling, wrong API usage, and async mistakes, plus a 5-bug hold-out set used only for final evaluation. A Creator agent proposes code patches using Qwen2.5-Coder-7B served through vLLM on AMD AI Notebooks with ROCm. A deterministic Judge applies each patch in a temp sandbox and accepts it only if the project builds and the Vitest suite passes. Every attempt is logged as a trajectory, so wins and failures become evidence. Between benchmark iterations, Malatang uses Fireworks for a bounded reflection step that rewrites the Creator’s own strategy playbook from sandbox-verified results. The product is not a magical bug fixer; it is the harness, contracts, evidence trail, and measured improvement curve. Our project shows an original, reproducible way to evaluate self-improving agents using real GPU-backed inference, strict verification, frozen benchmarks, and transparent integrity reporting.

To participate in the hackathon, simply click Sign up with AMD. If you are not already a member of the AMD AI Developer Program (ADP), you will need to create an account. ADP members can access AI tutorials, experts, and community support. In addition, new members can claim a free month of DeepLearning.ai pro, $100 in AMD GPU cloud credits, and $50 in Fireworks AI API credits. All hackathon participants will receive on hackathon launch additional compute and API credits for their projects. Ready to build what's next on AMD? Sign up with AMD below to secure your spot.