ShieldOn

Inspiration

Business investigations often pull from internal records and current external context. The hard part is keeping each claim traceable and making sure a recommendation does not quietly become an action. ShieldOn follows a simple rule: AI can investigate and recommend, but a human must decide.

What it does

A reviewer defines an investigation and adds up to six internal evidence items. ShieldOn uses Tavily to find current external sources related to that case, then sends a numbered evidence register to NVIDIA Nemotron 3 Super 120B through Nebius Token Factory. The model returns candidate findings tied to evidence IDs. Reviewers can inspect the sources, approve or reject each finding, and record a rationale. Only an explicitly approved finding can create a governed action. The decision register can then be exported for audit.

How we built it

The challenge edition uses Next.js 16 and TypeScript. Nebius Token Factory serves Nemotron through its OpenAI-compatible inference API. Tavily provides live, case-specific research, with source URLs and excerpts visible to the reviewer. The server validates model output, checks that cited evidence IDs exist, and keeps each finding pending until a human reviews it.

What we learned

Structured model responses can vary in their JSON envelope and field names. We added normalization and validation instead of trusting a single response shape. We also enforce the approval boundary in application logic and tests, rather than relying on prompt instructions. Tavily is most useful when its results remain visible in the evidence register, so reviewers can see what influenced a finding.

What changed during the hackathon

ShieldOn began as a broader product direction. For this challenge, we built a focused, working investigation application with custom evidence intake, dynamic Tavily research, Nebius-hosted Nemotron analysis, evidence-linked candidate findings, human approval and rejection, governed action generation, an exportable decision register, automated tests, and a public deployment. The repository documents the challenge-specific scope. The demo uses sample business data.

Built With

• factory
• nebius
• nemotron
• next.js
• nvidia
• tavily
• token
• typescript