docs/submission_notes.md

Submission Notes

Project Title Options

• RouterCore: Safe Routing for Agentic DevOps
• RouterCore: Fine-Tuned Workflow Routing Before Agent Execution
• RouterCore: Safe Routing, Not Just Routing
• RouterCore: Policy-Aware Routing for DevOps Agents

Short Description

RouterCore is a safe routing layer for agentic DevOps workflows. It converts natural-language requests into validated workflow routes, structured parameters, clarification questions, confirmation decisions, or policy rejections before an orchestrator can act.

Long Description

RouterCore demonstrates the step before agent execution: deciding whether a user request should be routed, clarified, confirmed, rejected, or sent to fallback. The MVP includes a deterministic router, schema validator, policy engine, iterative clarification state, and execution-preview orchestrator. It intentionally does not execute real cloud or infrastructure actions.

For the AMD Developer Hackathon, RouterCore is framed as a Track 2 fine-tuning project. It includes synthetic train/eval data, FakeRouter baseline evaluation, optional prompted Hugging Face model evaluation, and a LoRA fine-tuning path designed for AMD Developer Cloud with ROCm. A safety-tuned compact Qwen LoRA run on an AMD Instinct MI300X VM improved required-field presence from 28.57% to 100.00% and status accuracy from 57.33% to 86.67%, while preserving 100.00% unsafe rejection accuracy and 0.00% false route rate. The goal is to improve structured router proposals while preserving validation and policy redundancy.

Suggested Tags

• AI agents
• Fine-tuning
• AMD ROCm
• Hugging Face
• LoRA
• DevOps
• Safety
• Workflow routing
• Gradio
• Pydantic

Track Selection Recommendation

Submit under Track 2: Fine-Tuning on AMD GPUs.

RouterCore also has a Track 1-style agent workflow demo, but the strongest judging story is the fine-tuning objective: improve structured routing quality while preserving safe-routing behavior.

What To Submit On lablab

• Public GitHub repository
• Hugging Face Space link for the Gradio demo
• Short demo video
• README with setup instructions
• Baseline evaluation report
• Evaluation comparison report
• Notes on the LoRA fine-tuning run and AMD ROCm results

What To Include In The Video

• The problem: agents need a safe routing decision before execution.
• The app flow: router, validator, policy, orchestrator preview.
• A successful web app route.
• A missing-field clarification.
• A risky IAM owner request rejected by policy.
• Baseline metrics and what they imply.
• Fine-tuning result on AMD Developer Cloud with ROCm.

What To Include In The Slide Deck

• One-slide problem statement
• Architecture diagram
• Router output contract
• Policy redundancy examples
• Demo screenshots
• Baseline metrics table
• AMD LoRA fine-tuning result and eval comparison
• Next steps

Hugging Face Space Deployment Notes

Use the Gradio app as the Space entrypoint. The Space can run the deterministic FakeRouter by default so it stays lightweight and reliable.

Recommended Space files:

• app/gradio_app.py or a root-level app.py wrapper
• requirements.txt
• routercore/
• data/schemas/
• data/examples/

Keep prompted model and LoRA evaluation optional. Large model downloads should not be required for the public demo Space unless the Space hardware and storage are explicitly configured for it.