GroundTruth - Crisis Verification Agent for Slack

Inspiration

During disasters, the deadliest thing in a coordination channel isn't slow information — it's wrong information. "The bridge is open." "This water is safe." "Shelter at location X" (which flooded an hour ago). I've been co-authoring research on classifying safe vs. unsafe instructions in safety-critical text using RAG grounding, and I realized every Slack agent in existence answers, summarizes, or coordinates — none of them verify. GroundTruth is that missing safety layer.

What it does

GroundTruth monitors a crisis-coordination channel and checks every actionable claim against a corpus of official disaster advisories (NDMA / WHO / state authority guidance) using retrieval-augmented verification. Each claim gets a threaded verdict:

• ✅ Verified — with a citation to the official advisory
• ⚠️ Unverified — no official source found; treat with caution
• 🚨 Dangerous — contradicts official guidance; auto-escalated to #crisis-alerts with the grounded correction

A /groundtruth check <text> slash command lets anyone verify guidance on demand, and every verdict is written to an append-only audit log so responders can prove what was known, when.

How we built it

• Slack Agent Builder for the agent scaffold and channel event handling
• Slack AI capabilities for claim extraction from natural conversation
• MCP server integration — the verification engine runs as an MCP server: sentence-transformer embeddings + FAISS retrieval over the advisory corpus, with a verdict layer on top
• Block Kit for clean threaded verdict cards with severity, citation, and confidence

Challenges we ran into

Distinguishing actionable instructions from ordinary chatter — verifying every message would make the agent unbearably noisy. We solved it with a claim-extraction pass before verification, so GroundTruth only speaks when a message could cause someone to act.

Accomplishments we're proud of

A working end-to-end verify-and-escalate loop in under 18 hours, grounded in real published advisories — and a design where the agent stays silent unless it has something evidence-backed to say.

What we learned

Trust is a feature. An agent that says "I can't verify this" honestly is more valuable in a crisis than one that always has an answer.

What's next for GroundTruth

Multi-language claim checking (Hindi/Gujarati), live advisory ingestion via the Real-Time Search API, and piloting with a campus disaster-preparedness cell.

Built With

• block-kit
• faiss
• mcp
• natural-language-processing
• python
• rag
• sentence-transformers
• slack
• slack-agent-builder