ProofStack AI

Production-Quality AI Research and Evidence Intelligence Platform

Built for the Nebius x NVIDIA Global AI Hackathon

"Intelligence is easy. Evidence is everything."

ProofStack AI investigates complex questions, retrieves real-world web evidence, extracts factual claims, rigorously evaluates evidence support and contradictions, identifies boundaries of uncertainty, and generates transparent, source-backed research reports.


1. System Architecture

User Interface (React + TypeScript + Tailwind CSS)
    ↓
Research Orchestrator (FastAPI / Express Backend)
    ├── 1. Research Decomposition (Breaks down query into empirical targets)
    ├── 2. Evidence Retrieval (Tavily Search API)
    ├── 3. Claim Extraction (NVIDIA Nemotron via Nebius Token Factory)
    ├── 4. Evidence Assessment (SUPPORTED | CONTRADICTED | MIXED | INSUFFICIENT EVIDENCE)
    ├── 5. Contradiction Analysis (Exposes conflicting cross-source claims)
    └── 6. Uncertainty & Epistemic Boundaries Report
    ↓
Validated Research Report & Source Explorer
    ↓
Interactive Research Dashboard

2. Evidence Assessment Philosophy

A citation is never treated as proof merely because a document was retrieved. ProofStack AI enforces strict separation between:

  1. Finding a source (Relevance).
  2. Extracting an atomic factual claim (Passage attribution).
  3. Evaluating whether the passage genuinely supports, refutes, or is ambiguous toward the claim.
  4. Distinguishing empirical retrieved facts from model synthesis and recommendations.

Four Formal Evidence Labels:

  • SUPPORTED: The available source passages directly and unequivocally substantiate the claim.
  • CONTRADICTED: The available evidence directly refutes the claim with verifiable opposing findings.
  • MIXED: Materially conflicting evidence exists across peer sources or testing methodologies.
  • INSUFFICIENT EVIDENCE: The retrieved material does not establish the claim either way. Lack of evidence is never presented as proof that a claim is false.

3. Technology Stack

  • Frontend: React 19, TypeScript, Tailwind CSS v4, Lucide Icons, Motion.
  • AI Model: NVIDIA Nemotron (nvidia/llama-3.1-nemotron-70b-instruct) via Nebius Token Factory.
  • Web Search & Evidence Retrieval: Tavily Search API.
  • Backend:
    • Full-stack runtime via Express + Vite middleware (runs on port 3000 in unified container).
    • Standalone Python FastAPI service located in /backend with Pydantic validation schemas.
  • Database: In-memory session store with PostgreSQL adapter configuration (DATABASE_URL).

4. Setup & Configuration

A. Environment Variables (.env)

Copy .env.example to .env and configure your credentials:

# NEBIUS TOKEN FACTORY (NVIDIA NEMOTRON)
NEBIUS_API_KEY="your_nebius_api_key_here"
NEBIUS_BASE_URL="https://api.tokenfactory.nebius.ai/v1"
NEBIUS_MODEL="nvidia/llama-3.1-nemotron-70b-instruct"

# TAVILY WEB RESEARCH API
TAVILY_API_KEY="tvly-your_tavily_api_key_here"

# PERSISTENT DATABASE (Optional PostgreSQL)
DATABASE_URL="postgresql://user:password@localhost:5432/proofstack"

B. Running the Web Application

npm install
npm run dev
# The full application is accessible at http://localhost:3000

C. Running Standalone FastAPI Backend (Optional)

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

5. Pre-Indexed Benchmark Mode (Zero-Failure Guarantee)

If API keys are not supplied during evaluation or network constraints arise:

  • ProofStack AI provides an explicit, transparent Pre-Indexed Benchmark Mode using deterministic empirical research cases (such as SWE-bench open model comparisons and speculative decoding mathematical evaluations).
  • All benchmark data is visibly labeled as DEMO MODE (PRE-INDEXED BENCHMARK) with genuine paper citations and verified benchmark scores.
  • Live investigations run automatically as soon as Nebius and Tavily keys are configured.

6. License & Disclaimer

ProofStack AI is an open research tool. Retrieved quotes and summaries represent the content of external third-party sources.

Built With

Share this project:

Updates