TechAssembly

Prospect Research turns a small business owner's plain-words description of what they sell into a working outreach campaign: researched buyers, sourced evidence, and a message for each one that the owner approves. I built it for my own business. I run 504 Yard Grills & Drinks in Lagos. Selling corporate lunch packs and event catering meant cold calling offices and guessing who might buy, and marketing agencies wanted retainers we could not afford. It was built during the hackathon, between Oct 27 and Nov 4, 2026. How it works. You describe your business by typing or speaking. A Meta Muse Spark agent running on Cloudflare Workers plans searches for your area, opens each business's own web pages, and keeps only claims it can link to a source. It scores every prospect on fit, problem, readiness, reachability and confidence. The shortlist shows what was found and why, and every claim links to the page it came from. When the evidence gives no reason to contact someone now, it says "no timing signal found yet" instead of inventing one. Businesses that do not fit are set aside with a written reason, never silently dropped. The core promise is timing: know who will buy before you say hello. A venue advertising December dates, an office hiring dispatch riders, a hotel promoting its event halls. The message opens with that real fact, not a template. Messages are written per prospect, with WhatsApp, email and direct message versions and an angle that names the real trade between the owner and the prospect: a venue gets live event grilling for the weddings it hosts, an office gets grill packs for staff lunch days. The prompt forbids claiming services the owner does not offer, and a draft that cannot cite the prospect's own evidence is rejected in code and never shown. The owner can make any message warmer or more direct, or ask for another, then approves it and sends by hand. WhatsApp opens with the text pre-filled and nothing is sent unless a human presses send. Research runs on the server, so the owner can leave and come back while it works. Everything is scoped to the signed-in user (profiles, runs, prospects, drafts, campaigns and sender accounts), enforced server-side and covered by tests. Built with: Meta Model API (Muse Spark) for the reasoning steps and Muse Voice Transcribe for speaking the business description. Stack: Cloudflare Workers, Workflows and D1. Live demo: https://prospect-research.sam-5b1.workers.dev