Digi4 Prospector — B2B Lead Sourcing Engine

At Digi4, a Moroccan software agency targeting the French market, sales prospecting ran on time-consuming manual scouting: spotting companies with a genuine digital gap (an aging website, an outdated tech stack, no CRM) meant hours scattered across directories, social networks and word of mouth, with no technical check ever confirming the diagnosis before the first contact. Saad built Digi4 Prospector to replace that guesswork with a measurable signal: continuously detect companies showing a real digital gap, then audit them first-hand with genuine technical tools before a salesperson ever reaches out.

The core of the system is a six-stage pipeline, each stage independent and replayable:

• Sourcing — over 36 scheduled harvester modules pulling from the French SIRENE registry, France Travail job postings, tender/legal-notice registries (BOAMP, BODACC, TED), and international registries (UK Companies House, Norway's BRREG, GLEIF)
• Normalizer — drops legally-protected records, deduplicates, disqualifies poor fits
• Enrichment — contact/email discovery with mail-server validation
• Audit Engine — a genuine technical audit via Playwright and Lighthouse
• Prioritizer — a weighted heat score
• AI Brief Generator — a channel-agnostic, actionable sales brief

A human then picks the outreach channel — cold email, a warm partner introduction, a marketplace proposal, or an inbound reply — with cold email deliberately ranked last.

An internal audit caught a bug before launch that would have silently broken the tool: "signal" is a reserved keyword in MySQL/MariaDB, which corrupted the core leads table the first time the schema actually ran — fixed by backtick-quoting the name across the schema and models. The database deliberately carries no foreign key constraints; integrity is enforced in application code through a shared Repository base class, parent-existence pre-checks, transactional cascades, and a weekly orphan-sweeper cron job. The team also refuses to automate cold email until a manual pilot validates the channel — a deliberate anti-premature-automation discipline.

The backend runs on Node.js/Express 4 with JWT authentication via Passport, serving an Angular 19 frontend styled with Tailwind 3, backed by MySQL 8 (database digi4_prospector). The audit engine combines Playwright, Lighthouse, the Google PageSpeed Insights API and the Chrome UX Report for real performance measurements rather than guesses. The AI layer is provider-swappable — Gemini as the free-tier default, with fallback to OpenAI, Anthropic, Groq, Cerebras, OpenRouter or Mistral — and systematically strips personal data before any call to the free-tier API, re-inserting it locally in Node once the response comes back. Built across 29 commits over roughly four weeks.