-
-
Multi-dimensional threat matrix evaluating CVE exposure, architectural debt, licensing, and maintainer health.
-
Verified Grounding Citations: Live web sources and documentation retrieved via Tavily Search API with confidence scores.
-
VeriFact Sentinel terminal for tech due-diligence powered by NVIDIA Nemotron on Nebius Token Factory and Tavily Search.
-
Completed DAG pipeline and official recommendation verdict card providing actionable production deployment guidance.
-
Actionable recommendation verdict card and CVSS-aligned risk meter providing clear production advice.
-
Multi-agent DAG pipeline decomposing queries and streaming live web intelligence via Tavily API in real time.
-
Official executive due-diligence dossier with full technical specifications and verified live citations.
Inspiration
Every day, developers and companies want to use new AI libraries and open-source tools. But they always face big risks: Are there hidden security bugs? Is the project abandoned? Does the license allow commercial use?
When you ask standard AI chatbots, they often hallucinate or use outdated information. We built VeriFact Sentinel to give developers an autonomous detective that checks the real facts in seconds before they install risky software.
What it does
VeriFact Sentinel is an autonomous AI due-diligence agent:
- Plans Research: Breaks down your target (like LangChain, Supabase, or DeepSeek) into technical questions.
- Searches Live Web: Uses Tavily AI Search to find the latest documentation, security advisories, and CVE vulnerabilities.
- Stops Hallucinations: Uses NVIDIA Nemotron-70B on Nebius Token Factory to double-check facts and filter out fake marketing hype.
- Calculates Risk (0-100): Gives an easy-to-read Risk Meter (Green = Safe, Yellow = Caution, Red = Dangerous).
- Gives Clear Advice: Tells you directly: "Recommended for Production" or "Strictly Do Not Use" with simple, clear reasons why.
How we built it
- AI Brain: NVIDIA Nemotron-70B model hosted on high-speed Nebius Token Factory GPUs.
- Live Search: Tavily AI Search API for deep, real-time web grounding and source citations.
- Backend: Python and FastAPI with live event streaming so you can watch the agent think in real time.
- Frontend: Modern, dark-mode web dashboard built with Tailwind CSS.
- Testing & Deployment: Tested with automated Python tests (100% pass rate) and deployed live on cloud via Docker.
Challenges we ran into
Calibrating the risk scoring so that trusted technologies (like React or PostgreSQL) get safe low scores, while dangerous or vulnerable tools (like unpatched Log4j) get clear critical alerts based on real evidence.
Accomplishments that we're proud of
- Making the system fast: you get a full verified report in under 15 seconds.
- Providing direct human advice: users do not have to guess; the agent clearly explains whether to use or avoid the tool.
- Zero hallucinations: every single claim is backed by real web citations.
What we learned
We learned that combining open-weight models like NVIDIA Nemotron on Nebius Token Factory with live web search from Tavily creates an unbeatable combination for reliable AI agents.
What's next for VeriFact Sentinel
Creating an automatic GitHub bot that comments directly on code Pull Requests to warn developers before merging unsafe packages.
