Bridge Sourcing

The problem

Most EU buyers sourcing from MENA hit the same wall: supplier discovery is manual, qualification is gut-feel, and outreach hits dead Hotmail addresses from 2017. Meanwhile Egypt has a free-trade lane to the EU (0% tariff on most categories) that almost no one outside the country can navigate. The opportunity isn’t the AI; it’s the trade-route arbitrage. AI is just what makes the pipeline cheap enough to run continuously.

The pipeline — 3 agents

• Discovery agent. Crawls LinkedIn + Hunter + targeted Google search for suppliers in a category. Outputs a normalised supplier row with company, contact, category, scale signal.
• Qualification agent. Reads the supplier’s site + LinkedIn + shipping records via public manifests. Scores fit-to-buyer-spec on a 0–100 rubric with explainable reasoning.
• Outreach agent. Drafts a personalised first email per supplier in the buyer’s voice, scheduled via Zoho with deliverability checks (DKIM verified Apr 2026). Never auto-sends — the buyer reviews + sends.

The 82→96% scrape accuracy curve

Day one, the discovery agent extracted supplier fields at 82% accuracy on a held-out test set. Three months later, 96%. The 14-percentage-point gain came from eight specific changes:

1. Switched extraction model from JSON-mode-flagship to JSON-mode-mini with a stricter schema. Smaller model, tighter prompt, +3pp.
2. Added a separate validation pass with a different model. The “does this look right?” second opinion catches hallucinations the extractor didn’t flag. +2pp.
3. Pre-cleaned HTML before extraction. Stripped nav, footer, ads. Removes 40% of input tokens and 70% of nuisance text. +2pp.
4. Per-field confidence scoring. Low-confidence fields get a retry with a different prompt before being marked “unknown” instead of guessed. +2pp.
5. Country-specific date and address parsers. Egyptian addresses don’t match Western patterns. Hand-rolled regex, not the model. +1pp.
6. Duplicate detection across discovery batches. The same supplier appearing twice with different field values used to silently corrupt scoring. Now it triggers a merge step. +1pp.
7. Field-level golden eval set. Built a 200-row hand-labeled set and ran every prompt change against it before deploying. Stops regressions cold. +2pp.
8. Stopped trusting the model on numeric fields. Years founded, employee count, revenue band — all parsed by hand from canonical fields, not inferred from prose. +1pp.

What I got wrong