What NIGHTEYE does
NIGHTEYE Evidence Engine is an evidence-first Nemotron agent for people who need to know what materially changed without drowning in noise.
A user gives NIGHTEYE a monitoring question and a packet of timestamped evidence. The system:
- separates observation from inference;
- decides whether a material change occurred;
- generates competing explanations;
- names a falsifier for the leading explanation;
- preserves explicit unknowns;
- recommends reversible next actions;
- returns evidence IDs so the reasoning can be audited.
Technological implementation
The hackathon implementation lives on the public nebius-nighteye branch.
The reasoning path is wired to Nebius Token Factory through its OpenAI-compatible API and is configured to use NVIDIA Nemotron 3 Super 120B A12B by default.
Application stack:
- Python 3.12
- FastAPI
- Nebius Token Factory client
- NVIDIA Nemotron model path
- browser UI
- Docker
- deterministic evidence validation
- GitHub Actions
- dependency auditing
- Bandit
- CodeQL
NIGHTEYE treats supplied evidence text as untrusted data rather than instructions. After inference, it validates evidence references in both observations and competing explanations. Unsupported evidence IDs are rejected instead of silently treated as support.
The adversarial suite also checks malformed JSON, invalid material-change types, duplicate explanations, and prompt-injection-style evidence. Functional and security CI pass.
Product design
The browser experience is intentionally structured around one question: “Has something materially changed?”
Instead of returning one opaque answer, NIGHTEYE exposes the evidence-backed observations, competing explanations, falsifier, remaining unknowns, and safest next actions as separate parts of the brief.
The API key stays server-side.
Potential impact
NIGHTEYE is aimed at operators, analysts, founders, researchers, and teams that need to decide whether new information should actually change a decision.
Ordinary AI summaries often compress evidence and inference together. NIGHTEYE is designed to make unsupported confidence easier to detect and uncertainty easier to act on.
Why the idea is different
Nemotron is not used only to summarize. The surrounding application forces a falsifiable reasoning structure and then validates the model output before accepting it.
Track
Best Apps and Agents
Current verification status
Verified now:
- deterministic and adversarial tests pass;
- browser/API application is implemented;
- Docker packaging is present;
- dependency and security review passes;
- Token Factory/Nemotron client path is implemented.
Not yet claimed:
- successful live Nebius/Nemotron inference;
- public hosted demo;
- live-use quality ratings.
Those remain release gates until directly evidenced.
Significant update disclosure
This repository existed before the hackathon. The nebius-nighteye branch and nebius_nighteye/ application are the hackathon-specific implementation added during the submission window. Older Lead Rescue code is pre-existing and is not represented as new Nebius work.
Build credit
Entrant: Jose Padilla
AI engineering assistance: ChatGPT by OpenAI
Log in or sign up for Devpost to join the conversation.