Mojeeb Titilayo

Case study 04

Revel

An AI product strategist. Paste a URL and get a scored Reveal Index, a prioritised Blueprint and an exportable Action Queue, for people and for paying agents.

Studio product · BlindspotLabLiveJuly 2026

Problem

Founders can't see why a page isn't converting. Generic AI reviews say "improve your messaging" without quoting a single line of the page.

Product direction

Ground every finding in the page itself, make the output something you can act on (scored, prioritised, exportable), and make the same capability callable by AI agents, who pay per audit.

My role

Author and engineer: product, agent loop, provider cascade, MCP and payment integration. Prompt engineering by Mojeeb Titilayo, iterated with Claude and Grok Build CLI (credited in the repository's stacks.md).

Engineering approach

  • A custom single-agent tool loop (not LangChain) with three tools: fetch_url, discover_internal_links and submit_analysis_report. Cheerio extracts titles, headings, CTAs, navigation and body text as evidence.
  • A multi-provider cascade, Groq → OpenRouter → Gemini, falls through on recoverable failures (quota, incomplete reports, API errors). Zod quality gates reject reports that don't validate.
  • Site-grounding prompts require quoted page copy and ban generic advice. Provider and model errors are sanitised before they reach the UI.
  • A streamable HTTP MCP server at /api/mcp, with x402 payment gates only on billable tools, settled in USDT0 on X Layer. A Partner API issues keyed access.

Stack and constraints

Prompt engineering
Site-grounding promptsZod quality gatesMojeeb Titilayo (author)
Models
Groq llama-3.3-70b-versatileOpenRouter (env-pinned)Gemini 2.5 Flash / Flash Lite
Agents
Custom tool loopMCP (A2MCP)Partner API
Payments
OKX x402X Layer (eip155:196)
Product
Next.js 16React 19Supabase PostgresPrisma 7NextAuth v5Vercel
  • Free-tier model quotas, so a cascade was needed for reliability.
  • Long-running audits, up to 300 s per function.
  • Agent callers that pay per call: discovery must stay free and billing must be precise.

Design decisions

Cascade over a single provider

Ordering providers by speed and quota (fast free tier first, strongest grounded JSON second, tightest quota last) keeps audits finishing without exposing which provider failed.

Pay only for value

MCP discovery methods are free; x402 applies only to billable audit tools, matching the OKX agent marketplace payment path.

Quote or it didn't happen

Grounding prompts plus schema gates are what separate a usable audit from a generic one.

System

failfailURL or MCP callx402 on billable toolsTool loopfetch · discover · submitGroqllama-3.3-70bOpenRouterenv-pinnedGemini2.5 Flash · LiteZod gategrounded JSONReveal IndexBlueprint · Action Queue
Provider cascade with quality gates. Recoverable failures fall through; nothing unvalidated reaches a person or a paying agent.

Outcome and status

  • Live at tryrevel.xyz, with docs at docs.tryrevel.xyz.
  • Listed on the OKX AI marketplace as agent #4750, with its endpoint at tryrevel.xyz/api/mcp.
  • 82 commits between 2026-07-04 and 2026-07-27 in a public repository.

Sources: README and stacks.md