SameDayDesk Evidence Miner

Live x402 and Gemini evidence infrastructure that turns volatile web sources into cited, deterministic briefs for agents and small businesses.

Comment

Inspiration

Small operators and AI agents make decisions from web pages that change without notice and mix evidence with marketing claims. Search snippets are too shallow, while one-off manual research is slow and hard to audit. We wanted a machine-readable evidence layer that stays useful after the first answer.

What it does

SameDayDesk Evidence Miner accepts a URL, repository, company, or wallet target and returns structured evidence: clean text, metadata, JSON-LD, headings, AI-readiness signals, enrichment, security findings, and deterministic schemas. Seven production resources are exposed over HTTP and MCP, with pay-per-call USDC settlement on Base through x402.

A separate live Gemini service now reasons over supplied evidence excerpts, identifies supported claims and caveats, and returns a cited next-action brief. Deterministic application code rejects any claim that cites an evidence ID outside the supplied set, so users can inspect why a recommendation was retained.

How we built it

The production base is a Node.js and Express service using x402 protocol packages, the Model Context Protocol SDK, and Zod schemas. It runs on Railway, publishes a seven-resource discovery manifest, and returns HTTP 402 payment requirements before paid resources.

The XPRIZE layer is another bounded Node.js service on Google Cloud Run. It uses the official @google/genai SDK, Vertex AI Application Default Credentials, Gemini 2.5 Flash structured JSON output, a dedicated runtime identity, a fixed rate-limited public demo, and a private-key gate for arbitrary synthesis. Revision samedaydesk-evidence-miner-00003-fv9 serves 100 percent of traffic with min instances 0 and max instances 1.

Challenges we ran into

Payment middleware and network libraries change quickly, healthy RPC endpoints can still fail their first contract read, and live websites often omit the taxonomy fields their APIs promise. We learned to probe real contract reads, derive missing signals from observable content, and keep deterministic fallbacks.

The first live Gemini call exposed a more useful failure: a 900-token ceiling truncated otherwise schema-shaped JSON. We reduced the output shape, raised the bounded ceiling to 2,048 tokens, added finish diagnostics that never log model content, and redeployed. We also found that Cloud Run's frontend intercepted exact /healthz while /healthz/ reached Express, so the production canary moved to the unambiguous /readyz path.

Accomplishments that we're proud of

  • A live seven-resource x402 gateway on Base
  • HTTP and remote MCP access with a public discovery document
  • A credited Cloud Run service using Gemini 2.5 Flash through Vertex AI
  • A live demo returning three verified claims, zero unsupported citations, and grounding passed
  • Deterministic schemas and 13 passing regression tests across both services
  • Public MIT-licensed source and a verified Smithery listing
  • Cost controls: min 0, max 1, protected arbitrary synthesis, and a 25 USD monthly gross budget

What we learned

Generative reasoning is most valuable above a reliable evidence substrate. Deterministic extraction catches fabricated or missing fields; Gemini can connect supplied sources and explain uncertainty. Keeping both layers makes the result more useful to customers and easier to evaluate. A successful deployment is also not product-market fit, so we report zero users and unattributed receipts honestly.

What's next

The public sub-three-minute demo is now live. We will add evidence-lineage exports and longer-running observability records, test adversarial evidence packets, onboard three arms-length pilot users, and measure attributable revenue, repeat use, time saved, latency, and per-run model cost.

Built With

Share this project:

Updates