Nick Noiceless@nickiriskin79871

veriroute is one container that detects its task from the input schema and runs the right agent — both built on the same harness philosophy: deterministic routing, code-verified answers, escalate only proven failures. Track 1 — token-efficient router. A local Qwen2.5-1.5B answers sentiment/NER/summarization behind format verifiers, math via program-of-thought (the model writes solve(), a sandbox executes it), codegen via generated self-tests. Only verified failures and factual recall escalate to the best non-thinking model in ALLOWED_MODELS. Guardrails: stub-first atomic output, ALLOWED_MODELS asserted before any network I/O, hard token budget, prompt-prefix prewarming to fit 2-vCPU windows. Measured on a grader-class VM: 4/8 practice tasks answered free, 2,305 tokens total. Track 2 — all-local captioner. ffmpeg frames -> SmolVLM2-500M describes with a cross-frame consistency check -> Gemma 3 4B writes all four caption styles in one few-shot call. Few-shot beat two LoRA fine-tunes (Fireworks SFT llama-8b and our own GPU LoRA) in a blind judged bake-off — we measured, then chose. All 12 example captions ship in genuine distinct styles at 259s/3 clips on worst-case hardware. 100 tests including SIGKILL-resilience; submission journal with prediction-vs-leaderboard tracking in the repo. R&D (dataset distillation, LoRA training runs, bake-off harness): github.com/bogdan-lmk/gemmacap.

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.