🌿 ORIGIN

Find the cause, not the symptom.

---

Most "AI agent" demos ask an LLM a question and print whatever it says back. That always bothered me — a model can sound completely certain about something it just made up. So ORIGIN doesn't work that way. You give it a claim about a place — "this site had no deforestation," "there's no protected area nearby," "this project reduced emissions by 40%" — and it refuses to just answer. It goes and checks: real satellite data, real government registries, real vegetation indices, actually pulled for the actual coordinates in your claim. Only after eleven agents come back with real findings does anything get synthesized into a verdict, and even then it's phrased as evidence for a human to weigh, not a verdict handed down from on high.

I built the whole thing solo, and most of the actual work went into resisting the urge to fake anything — no seeded data, no "trust me" scores, no black-box confidence numbers. If ORIGIN says a claim is contradicted, you can click through to the exact hectare count, the exact NDVI delta, the exact source it came from.

How it actually works

1. Understand the claim. Location Grounding pulls a place out of the claim text and resolves it to coordinates. Claim Decomposition splits compound claims ("reduced emissions by 40% and has no protected areas nearby") into pieces that can be checked independently. These two run at the same time — neither needs the other's output.

2. Go find out. Seven agents fan out in parallel, each hitting a real data source: tree-cover loss from Global Forest Watch, protected-area proximity from WDPA, water/disaster risk from GDACS, a Gemini vision pass over actual satellite imagery, vegetation trend from Earth Engine's NDVI, carbon-offset project registries via Carbonmark, and long-run climate trend from NASA POWER. If a source has nothing to say, it says nothing — no fabricated "insufficient data" filler dressed up as a real result.

3. Weigh it, then decide. Cross-Reference goes through every finding and checks it against the claim's exact wording — this is the part that actually catches things, like noticing "no deforestation" plus a real tree-loss finding is a contradiction, not two unrelated facts. Verdict Synthesis then turns that into a direction score, a coverage score, and a written verdict with every citation intact. It won't tell you a claim is definitively true or false — it tells you what the evidence says and lets you decide.

If it can't pin down a location with real confidence, it stops and asks instead of guessing. It also won't try to resolve "Company X" into a specific facility — that's a different, harder problem, and bolting a bad guess onto it would undermine everything downstream that depends on the location being right.

What's actually real here

• Gemini, through Vertex AI, does the multimodal vision pass on satellite imagery, the evidence cross-referencing, verdict writing, claim splitting, and location extraction — every call uses Gemini's native structured-output mode, not a prompt begging for JSON and a regex cleaning up whatever came back.
• Google ADK composes the whole thing as a real SequentialAgent/ParallelAgent graph. The concurrency is real, too — Location Grounding and Claim Decomposition genuinely run at the same time, and all seven evidence agents genuinely fan out together the moment a location resolves.
• A2A carries every handoff between agents as a structured a2a.types.Message, not a dict I made up and called a protocol.
• MCP — a real FastMCP server (mcp_servers/origin_tools.py) hosts the data-fetching tools, and three of the evidence agents call into it as an actual subprocess, not a function call wearing an MCP costume.
• Earth Engine runs real NDVI queries against Sentinel-2 imagery.

None of the underlying data is fabricated or seeded: Global Forest Watch, WDPA, GDACS, Nominatim, Carbonmark, NASA POWER, and Esri World Imagery (the actual image Gemini looks at) are all live, real, and queried fresh per investigation.

Delivery is FastAPI over Server-Sent Events, so you watch each agent report back the moment it finishes instead of staring at a spinner.

Getting started

Prerequisites

• Python (developed and tested on 3.14)
• A Gemini API key — if you hit a limit: 0 quota wall on a plain AI Studio key, see agents/gemini_config.py. This routes through Vertex AI Express Mode instead, which has its own genuinely free quota pool.
• A Global Forest Watch account and API key — python3 scripts/gfw_create_key.py will generate one for you (your password stays local, typed via getpass, never leaves your terminal).

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure

cp .env.example .env
# fill in GEMINI_API_KEY and GFW_API_KEY

EARTH_ENGINE_PROJECT is optional — Vegetation Trend just reports "no signal" without it. To turn it on, register a Cloud project for Earth Engine's free noncommercial Community tier (console.cloud.google.com/earth-engine) and run earthengine authenticate once.

Run

uvicorn server.app:app --reload

Open http://localhost:8000. Investigations is where you submit a claim; Layers walks through the actual stack in more depth; Saved reports is everything you've investigated so far.

There's also a CLI if you just want one claim without the UI:

python3 -m orchestrator.pipeline "the palm oil expansion at -9.98,-63.0 in Rondonia, Brazil caused no deforestation"

Deploying it

render.yaml is already set up for Render's free tier — no card required to sign up. To deploy:

1. Push this repo to your own GitHub account (fork it, or just push to a repo you own).
2. On Render, choose New > Blueprint and point it at that repo. It'll pick up render.yaml automatically.
3. Render will ask for GEMINI_API_KEY and GFW_API_KEY — paste the same values from your .env.
4. Deploy. First load after the service has been idle takes 30-60 seconds to wake back up — that's Render's free tier sleeping after 15 minutes of no traffic, not a bug.

Two things worth knowing about the free tier specifically: EARTH_ENGINE_PROJECT is deliberately left unset in render.yaml, since Earth Engine's OAuth credential is a local file earthengine authenticate writes to your own machine, not something that travels with the repo — Vegetation Trend just reports "no signal" in the deployed version, the same graceful skip it does for anyone who hasn't run that command locally either. And the SQLite claims log resets on every new deploy (no persistent disk on the free tier), so Saved Reports only holds what's been investigated since the last push, not a permanent history.

Project structure

agents/            Gemini-backed reasoning agents (location grounding,
                    claim decomposition, visual inspection, cross-reference,
                    verdict synthesis)
data_clients/       Thin, real API clients per data source (GFW, WDPA,
                    GDACS, Earth Engine, Carbonmark, NASA POWER, Nominatim)
orchestrator/       ADK pipeline composition, A2A messaging, MCP client,
                    evidence/claims log
mcp_servers/        The MCP tool server the pipeline's evidence agents
                    call into
server/             FastAPI app + the frontend (vanilla HTML/CSS/JS, no
                    build step)
docs/brief.md       The original project brief and scope boundaries