NEXUS — Autonomous Intelligence Command Center

NEXUS turns messy real-world problems into evidence-backed action — investigating data, ranking competing hypotheses, simulating interventions, and learning from outcomes.

Comment

Inspiration

Community organizations constantly decide where limited resources should go — food, education, outreach, transport — while drowning in messy, incomplete, conflicting information. Chatbots give a single confident answer with no evidence trail. Dashboards show numbers but never recommend. We wanted an AI that investigates instead of answering: one that shows its work, admits uncertainty, and lets humans stay in charge.

What it does

NEXUS — Autonomous Intelligence Command Center turns a messy real-world problem into evidence-backed action through a visible loop:

Problem → Evidence → Hypotheses → Simulation → Action → Outcome → Learning

Give it a problem (our demo: community service prioritization across neighborhoods). NEXUS structures it, profiles the data with real statistics, ranks competing hypotheses with evidence-weighted scores, simulates interventions (Baseline vs A vs B — always labeled estimates), recommends an action a human must Accept / Modify / Reject, measures the recorded outcome, and learns. Every screen carries its receipts: an investigation timeline, a downloadable audit receipt, and honest labels everywhere (scores are not probabilities, simulations are not guarantees, demo data is synthetic).

How we built it

  • Frontend: React 18 + TypeScript + Vite + Tailwind CSS + Recharts + Lucide icons. Sidebar command center, ⌘K command palette, animated evidence graph (SVG), Scenario Lab charts, dark-first design system with light mode, reduced-motion support, keyboard navigation throughout.
  • Backend: Python + FastAPI + Pydantic + SQLAlchemy (PostgreSQL-compatible, SQLite locally). A real InvestigationOrchestrator runs Problem Parser → Evidence Analyzer → Hypothesis Generator/Ranker → Root-Cause Analyzer → Simulation Engine → Recommendation Engine → Feedback Engine. Every stage produces structured, validated output stored in the database.
  • AI: a provider abstraction (LLM_PROVIDER=demo|generic). Demo Mode is fully deterministic — pandas/NumPy analysis on a synthetic 180-row dataset — so the 3-minute demo never depends on a paid API. Live mode degrades safely to deterministic output offline.
  • Validation: 8/8 pytest tests, reproducible simulations (byte-identical re-runs), input validation with proper 422/400/404 handling, secret-free repo.

Challenges we ran into

  • Tailwind silently not compiling (missing PostCSS config) made the app look like raw HTML — diagnosed via the 2.8 KB CSS bundle, fixed the pipeline, rebuilt the entire design system.
  • Honesty vs impressiveness: resisting the urge to fake progress bars or present scores as probabilities. We made transparency the brand instead — judges respond to it.
  • One-click reliability: the demo had to complete in ~10 seconds with zero config, so every external dependency got a deterministic fallback.

Accomplishments that we're proud of

A genuinely working Evidence → Hypothesis → Simulation → Action → Feedback loop (not a mockup), a downloadable investigation receipt, human-in-the-loop review gates, and a demo that runs fully offline.

What we learned

That the hardest part of decision-intelligence UX is communicating uncertainty without losing trust — and that labeling estimates honestly increases credibility instead of reducing it.

What's next for NEXUS

Real LLM reasoning behind the provider abstraction, causal-inference upgrades (propensity scoring, difference-in-differences), auth + shared workspaces, scheduled re-evaluation with drift alerts, and PDF receipt export.

Built With

Share this project:

Updates

Submission history