Inspiration
At 22, I built my consulting firm in Dubai and initially focused on Conversion Rate Optimization (CRO) for eCommerce stores. A tedious and time consuming task was to manually replay all user journeys to find CRO opportunities. AI can completely automate that process, and bring it further.
DTC sites lose revenue every day from silent bugs — a broken Add-to-Cart on mobile, a price drift between Google's snippet and the PDP, a stockout that's still being promoted. None of this pages Datadog. Nobody gets woken up. Customers just bounce.
We wanted an agent that watches the funnel like a real shopper, and when it spots a leak, doesn't just alert — it ships the fix.
What it does
LeakHound continuously monitors a merchant's storefront for revenue-leak invariants (broken purchase path, out-of-stock promoted, price/offer mismatch, trust gap, performance regression, missing competitor parity). When a leak fires:
Verifies it with a second model (no hallucinated leaks) Scores the revenue impact Opens a draft PR with the actual code fix Files a Linear ticket, pings Slack, publishes a public incident page Sells the full evidence bundle for USDC via x402, so other agents can buy and act on it One agent, six surfaces, one source of truth.
Details: Discover — Datadog Synthetic browser test fails → webhook kicks the FSM. (Or scheduled scan.) Crawl in parallel — Nimble Extract on the merchant's PDPs; Nimble SERP + LLM-platform brand-mention agents (ChatGPT/Gemini/Perplexity/Grok) for competitor parity; Stagehand v3 / Playwright cascade on Browserbase for the actual checkout funnel. Score revenue impact via DSPy ChainOfThought severity scorer (with heuristic baseline fallback). Verify in two stages: deterministic programmatic invariants keyed by funnel state (LASER FSM), then a cross-model ensemble (Claude Sonnet 4.6 + GPT-5.4 with position-swap and length-normalization). Assemble evidence with Claude Opus 4.7 + Anthropic Citations API, with AST-level citation validation — every [cite:evidence_id] must resolve to a real ClickHouse row or publication is refused. Fan out a single typed LeakEvidence to six surfaces in parallel: Slack alert, GitHub draft PR with a real candidate-fix code diff, Linear issue, Datadog Event, public incident page (JSON-LD Report + Satori OG), and an x402-paywalled JSON endpoint (Base USDC via Coinbase CDP) for downstream agents. Plus an MCP server (streamable-HTTP) exposing 4 tools so other agents can natively query LeakHound: scan_preview, get_incident_preview, unlock_incident (pays the paywall and returns the full bundle), list_recent_incidents.
How we built it
Five small services that talk over typed JSON: a Next.js storefront, public incident pages, a TypeScript action layer (Hono), a Stagehand browser worker, and a Python LangGraph FSM. ClickHouse holds the evidence. Claude + GPT cross-check each other. Coinbase x402 + CDP handle the paywall.
The whole thing boots with make dev.
Challenges we ran into
Mostly: 2026 APIs don't match 2024 docs. Nimble changed its request shape. Anthropic dropped temperature. OpenAI's Responses API rejected response_format. Stagehand v3's page isn't a Playwright page. CDP renamed its wallet product. Datadog returns HTML on errors.
I also blew up the FSM a few times on parallel state writes before learning LangGraph's reducer contract, and accidentally collapsed 5 distinct leaks into 1 PR before fixing our slug to include the URL path.
Accomplishments that we're proud of
A real agent that opens real draft PRs with real code fixes Cross-model verification — we refuse to publish unless every citation resolves to a real evidence row A first-class MCP server, so other agents can use LeakHound as a tool An x402 paywall that actually settles in USDC on Base All docs and vendor-quirk rules captured for fast continuation of the project.
What we learned
Trust the SDK's types, not the docs. Build stubs for everything before you have keys. One typed contract is worth more than ten clever adapters. And when an agent publishes findings to the world, citation validation is a hard gate, not a nice-to-have.
What's next for LeakHound
Repo-fingerprinting so we can ship fixes to Shopify, Magento, Woo, BigCommerce — not just our demo A deeper investigation mode for high-severity leaks Listing on Coinbase Bazaar so other agents can discover and pay us Multi-tenant, per-merchant cost caps
Log in or sign up for Devpost to join the conversation.