Live Web Evidence Agent

Inspiration

AI can summarize the web quickly, but speed is not enough. Teams still need to know which claims came from which source, what confidence level each claim deserves, and which sentences should not be repeated until a human approves them.

Live Web Evidence Agent is built for that gap. It turns a research answer into a visible evidence workflow instead of a black-box paragraph.

What it does

The app turns a research question into:

source cards
extracted claims
confidence labels
blocked claims
follow-up search plan
voice handoff gate
approval-ready answer

The key product idea is that the answer is not just a paragraph. It is a claim ledger with a clear point where the agent calls a human back in.

How I built it

The current public prototype is a static HTML/CSS/JavaScript app. It includes a refresh script that fetches public DeveloperWeek pages and writes a dated evidence snapshot into data/evidence_snapshot.json. The browser app turns that snapshot into source cards and claim rows.

I also added a human handoff gate. When the agent has blocked claims, it prepares a short resume packet and a browser-side voice prompt. The agent can keep moving until judgment is required, then call a human back to review the exact risky claims.

The intended sponsor integration is Nimble. The next version can connect Nimble Search, Extract, Crawl, Map, Web Search Agents, MCP server, or APIs to populate the same evidence table with live web data.

Challenges

The hardest design choice was not to overclaim. The public static MVP does not claim real Nimble API usage yet. It demonstrates the workflow that live web data should feed, and keeps unsafe claims blocked until a human approves them.