Hackathon entry
Deal intelligence you can verify
An engine that discovers AI pricing deals with agents, verifies every claim with a WASM fact-checker, normalizes them into comparable math, and serves everything as static APIs for humans and machines.
28
verified offers
70
priced models
60
scored routes
64
quotes wasm-verified
8
agent tools via MCP
The pipeline
1 · DISCOVER
A hermes agent crawls official pricing pages only. Every target it resolves is remembered in a crawl ledger so proven dead ends are never chased again.
2 · VERIFY
Each evidence quote is checked verbatim against the live page, then scored semantically by FactJudge — a deterministic WASM NLI model evolved against adversarial benchmarks (99.2% heldout). Contradicted prices are refused at merge.
3 · NORMALIZE
Deterministic math converts every offer into comparable units: effective $/Mtok under an agent request profile, value multiples of subscription pools, mega-deal detection with stated reasons.
4 · PUBLISH
Static build emits the site, JSON APIs and markdown digests to the edge. No backend, no marginal cost per reader, git history is the change feed.
Query it like an agent
# daily digest for any LLM curl -s https://llmdeals-v2.pages.dev/api/v1/daily.md # full machine-readable DB curl -s https://llmdeals-v2.pages.dev/api/v1/deals.json | jq '.deals[0]' # register the MCP server, then ask naturally: # "compare grok-4.3 vs glm-5.3 effective prices" # "fact_check: does this page support '$20/month'?"
Machine endpoints
| Method | Endpoint | Returns |
|---|---|---|
| GET | /api/v1/daily.md | daily markdown digest — top deals, what moved |
| GET | /api/v1/deals.json | full DB: 60 scored routes, evidence quotes, FactJudge verdicts |
| GET | /api/v1/answers/index.md | pre-rendered agent answers with live web context |
| MCP | mcp/server.mjs | 8 tools incl. fact_check() — run our WASM verifier on any claim |
Honest limitations
- Machine-discovered deals publish as unrated until editorial review — the homepage board only carries human-ranked verdicts.
- FactJudge scores semantic support, not absolute truth; thresholds (0.3 / 0.6) are calibrated but a judge is not a source. Every score links back to the official URL.
- JS-walled pricing pages can't be re-verified by plain fetches — those claims are marked
uncovered, never silently trusted. - Negative findings are kept: two researched providers had no real deal — they live in the ledger so no one re-chases them.
Built on
Astro static site · Cloudflare Pages · hermes agent (opencode-go) · FactJudge c026port WASM · SerpApi (guarded to free tier) · zero-backend JSON/markdown APIs. Everything runs on a 4-core box for $0/month.