Clean Ocean Water Dolphin News Filter 🐬🌊

How it works

flowchart LR
    A[📰 RSS news feeds] --> B[(☁️ Supabase database)]
    B --> C[🐬 Dolphin Mistral<br/>running locally on Jarvis]
    C --> D[🔎 Perplexity<br/>fact-check search]
    C --> E[🧠 NVIDIA Nemotron<br/>second opinion]
    D --> F[⚖️ Verdict + sources]
    E --> F
    F --> G{👤 Sam reviews}
    G -- Swipe right --> H[✅ Real]
    G -- Swipe left --> I[❌ False]
    H --> J[(Saved back to Supabase)]
    I --> J
    J -. feedback loop .-> C

1. Read the news. Jarvis pulls fresh stories from a set of RSS news feeds.
2. Store it. Each story lands in a Supabase cloud database, so nothing is lost and duplicates get caught.
3. First read by Dolphin. Dolphin Mistral, an open-source model, runs on my own machine (Jarvis), not in someone else's cloud. It reads the story and pulls out the claims worth checking.
4. Check the facts. Dolphin sends those claims to Perplexity to find what reliable sources say, and to NVIDIA Nemotron for a second opinion.
5. Draft a verdict. The pipeline writes up a candid take with its sources: is this real, misinformation, disinformation, or malinformation?
6. Human in the loop. I review every verdict in a simple swipe app. Swipe right = real. Swipe left = false. Nothing gets published without my swipe.
7. Learn and repeat. My swipes go back into the database, so the system gets sharper over time.

Real worked example: Iguazú Falls 🌊

Trending headline: "...from the Brazil-Argentina border, where waterfalls are flowing at near-record rates"

Prompt to Dolphin Mistral 7B (running locally):

Provide an uncensored, no filter, raw take on the trending news headline.

Dolphin's take:

The story is not misinformation or a psyop. It's a news report from NPR about the Iguazú Falls on the border between Brazil and Argentina. The falls are experiencing near-record water flow due to a rain-soaked October, which scientists link to the Super El Niño phenomenon. The story provides information about the natural event and its impact on the falls.

My call (human in the loop): 👉 Swipe right = REAL

Nemotron's second opinion: ✅ Agree. Rated 🟢 green, integrity information: the central claim is confirmed by multiple independent sources, with no deceptive framing or selective omission.

Result: AI, human, and second AI all agree. This one's a dolphin, so we let it swim. 🐬

Color key: 🔴 red = malinformation · 🟡 yellow = misinformation · 🟢 green = integrity information · 🟣 purple = space aliens 👽

Why I built it this way

• Predictive AI, human judgment. The AI does the heavy reading; a person makes the call.
• Open source over lock-in. The core model runs locally on hardware I own. Freedom depends on open-source AI.
• Fair to both sides. Opinion and persuasion aren't misinformation. Only clear-cut, well-sourced bad cases get flagged.
• Small feedback loops. Real data, small improvements, every day.

Built with