ForgeXi is an evidence-driven coding and agentic engineering system for the Nebius x NVIDIA Global AI Hackathon, targeting the Coding and Agentic Engineering track.

The system is built around a bounded Plan → Act → Test → Diagnose → Repair → Verify loop. NVIDIA Nemotron on Nebius is the reasoning layer; repository execution is isolated behind an allowlisted executor; authentication is capability-negotiated and kept outside trajectories; and each run produces tamper-evident verification evidence.

Implemented architecture includes capability-driven IAM/Token Factory authentication selection, short-lived IAM credential support for Nebius AI Cloud, a secret-free /models capability probe, normalized inference receipts containing model/latency/token counts/response digest, a bounded repair loop, paired benchmark structures, and five native WebMCP-facing domain operations.

ForgeXi deliberately does not claim benchmark wins, reliability, latency, cost, or live Nebius success until those measurements have been executed. The remaining submission gates are live Nemotron inference on Nebius, real repository benchmark runs, a hosted working demo, and the public ≤3 minute demo video.

Built With

  • github
  • nebius-ai-cloud
  • nebius-token-factory
  • nvidia-nemotron
  • pytest
  • python
  • webmcp
Share this project:

Updates