The context
Outbound sales to local businesses is mostly manual grunt work: scraping directories, eyeballing whether each business is worth calling, guessing what to say, and re-keying it all into a spreadsheet. Rian Infotech built an AI Lead Finder that chains hosted discovery, browser-impersonated website reading, objective site-quality scoring, and LLM judgment into one pipeline, then accumulates every result in a team CRM with role-based access and a per-action cost ledger.
The problem, precisely
Manual discovery is slow and risky. Finding businesses by hand means scraping directories one by one, which is fragile, against most sites' terms, and gets IPs blacklisted in minutes. The good leads get buried in the noise.
No fast way to judge fit. Reps eyeball each listing to guess whether a business is worth calling, with no consistent signal for whether it has a weak web presence, real revenue, or is just a national chain too big to sell to.
Guessing what to say on the call. Without a tailored opener, callers improvise a pitch for every lead, which is inconsistent and slow, and the reason a lead is worth a call is rarely written down anywhere.
Results scatter and spend is invisible. Lead data lands in ad-hoc spreadsheets, human notes get lost on re-runs, and the cost of each paid API call is never tracked, so nobody can audit what a run actually spent.
What we built ✓ verified in code
A multi-stage discovery and enrichment pipeline
One configurable pipeline turns a city plus a niche into outreach-ready leads: hosted Google Maps discovery via Apify, size and quality filters that drop dead listings and oversized chains, then browser-impersonated website reading (fast HTTP first, real Chromium fallback) to extract tech stack, social profiles, and public emails.
Explainable web-presence classification with a PageSpeed rescue
A deterministic rule engine labels each site none, basic, or good with a plain-language reason, reserving LLMs for the judgment calls. Google PageSpeed runs only on otherwise-good sites and downgrades slow ones back to basic, turning a discard into a speed-or-rebuild pitch.
LLM judgment layers for pain, fit, and people
Optional cheap-LLM passes mine negative reviews for recurring operational pain and a sellable fix, verify a lead actually matches the ideal customer profile, and look up the founder plus their LinkedIn URL, each with a deterministic fallback so the pipeline never hard-stops.
Priority scoring and an AI cold-call opener
A pure-function composite scores every lead 0-100 from website gap, review volume, rating, and other signals so callers work the best leads first, and an LLM writes one warm, spoken-sounding cold-call opener per lead that names the specific website gap.
A cost-audited team CRM with role-based access
Every result accumulates in a Supabase-backed CRM that dedups and merges across runs while preserving human notes and statuses. Three roles gate access, a styled Excel export mirrors the active filters, and an audit log records the exact USD cost of every money-spending action.
How it works
- 1
Discover. Hosted Google Maps scraping via Apify pulls businesses for one city and niche, then in-code filters keep active, premium operators with a phone number, deliberately keeping no-website businesses as the hottest leads and dropping chains too big to sell to.
- 2
Enrich. For each lead with a site, a fast browser-impersonated fetch (falling back to real Chromium on a block) reads the page to fingerprint its tech stack, support widgets, payment processors, social profiles, and public emails.
- 3
Classify and score. A deterministic rule engine labels web presence none, basic, or good with a reason; PageSpeed runs only on good sites and rescues slow ones to basic; then a 0-100 priority score ranks every lead so the best ones rise to the top.
- 4
Reason with LLMs. Cheap-LLM-first passes mine negative reviews for operational pain, verify ideal-customer fit, find the founder and LinkedIn, and write a tailored cold-call opener, each with a deterministic fallback so a missing key or failed call never stops the run.
- 5
Persist and audit. Leads are upserted into a Supabase or JSON CRM that merges across runs while keeping human notes, with a styled Excel export and an activity log that records who spent what, to the cent, on every paid action.
The outcome
The result is a repeatable, low-cost, auditable lead machine: every paid API call is tracked to the cent, every verdict shows its reasoning, human notes survive re-runs, and the team starts each day with the best leads, and the right opening line, already at the top of the list. A non-technical teammate runs it from a browser, watches a live progress bar, and downloads a polished Excel, or works the always-on CRM directly. Built on a resilient design with per-lead isolation, browser and heuristic fallbacks, checkpoint saves, and a run watchdog, so no single failure aborts a run.