Inspiration
Financial and trading systems operating with autonomous LLMs face a critical challenge: LLMs hallucinate prices, misinterpret market volatility, and generate unfounded trading recommendations. When high-stakes decisions depend on AI, fail-closed architectures are mandatory. VELQIRIS Nemotron Market Truth Guard was inspired by combining NVIDIA Nemotron's state-of-the-art reasoning and steerability with Nebius Cloud's high-performance AI inference infrastructure.
What it does
- NVIDIA Nemotron Integration: Leverages NVIDIA Nemotron reasoning models hosted on Nebius Token Factory for complex market regime analysis and risk classification.
- Fail-Closed Architecture: When API tokens, verifiable telemetry, or model IDs are unavailable, the agent fails closed, recording zero fabricated outputs and refusing to synthesize market data.
- Structured Telemetry Verification: Ingests real-time venue telemetry, cross-validates divergence thresholds, and feeds verified contexts into Nemotron prompts.
- Deterministic Output Logging: Redacted cryptographic proofs of prompt inputs and inference responses ensure auditable agent behavior.
How we built it
- Engine: Python 3.12 market risk reasoning client (nemotron_market_agent.py).
- Cloud Infrastructure: Nebius Token Factory API (OpenAI-compatible REST endpoint) with NVIDIA Nemotron LLMs.
- Public Repository: https://github.com/genesiscode2026/velqiris-nebius-nemotron-evidence
- Open Source License: MIT License with public test verification suites.
Challenges we ran into
Ensuring zero mock-data leakage in production-reachable code paths while testing against dynamic cloud endpoints required designing strict environment-driven gate barriers.
Accomplishments that we are proud of
- Clean, deterministic fail-closed execution model that guarantees verifiable provenance.
- Highly efficient token throughput leveraging Nebius Cloud's NVIDIA GPU clusters.
- 100% passing test assertions verifying that invalid or missing inputs never produce unverified outputs.
What we learned
Building enterprise-grade financial AI agents requires treating reasoning models as verifiable verification nodes rather than conversational chatbots.
What is next
Deploying continuous streaming inference on Nebius Cloud clusters to monitor high-frequency market order books with sub-100ms Nemotron reasoning latency.
Log in or sign up for Devpost to join the conversation.