A bespoke five-stage scraping, structuring, and classification pipeline — built from scratch for a market existing data platforms didn’t serve.
A small Israeli financial services firm needed targeted prospects. The available enterprise data platforms didn’t serve the market — wrong language, wrong geography, wrong price point.
Sales worked from purchased lead lists with no geographic or income filtering. Conversion was poor because the prospect base wasn’t the right one — cold-calling people who had no need for wealth management or pension planning.
Enterprise data platforms (ZoomInfo, Clearbit, Apollo) were US-focused, priced at enterprise tier, and had thin coverage of Israeli business and professional records. None of them indexed the Hebrew-language sources where the actual prospects lived.
Filtering prospects by income tier, neighborhood, or postal code — the dimensions that actually predicted whether someone needed financial services — required a custom layer that no off-the-shelf tool offered for this region.
Each stage was a standalone tool with a single job. Nothing was end-to-end automated — I ran, validated, and handed off each component myself. Outputs landed on the sales team’s shared drive in Excel.
Israeli postal codes and neighborhoods classified into income tiers using public demographic and real-estate signals. The geo-target list became the scope for every downstream stage — collapsing the search space before a single page was fetched.
WebScraper (open-source) handled local extraction from registries, directories, listings, and municipal sites. BrightData’s Residential Proxy provided rotating IPs and anti-bot bypass for rate-limited and geo-restricted sources.
Custom VBA Macro scripts in Excel handled normalization, enrichment, formatting, classification, and dedup — turning raw messy data into a structured, highly filterable index with multi-search and sort capabilities. Every record had a known shape and a unique identity before any score touched it.
Each record was scored against weighted signals — geo-tier, profession, business-ownership indicator, cross-source coverage — to produce a ranked prospect list. Weights were tuned manually between batches based on conversion-rate feedback from the sales team.
The final stage exported themed Excel workbooks — sorted by region, profession, and score band — to the sales team’s shared drive. The team never had to interact with the pipeline; they just opened a file in the tool they already used.
A pipeline is a series of filters, not a series of steps. Each stage shrinks the input for the next — geo collapses scope before a page is fetched, dedupe collapses records before a score is computed. That discipline is what kept a manually-orchestrated system tractable.
The scraping was the easy part.
The classifier was the product.
The lessons from a 17-month custom pipeline outlasted the pipeline itself. They recurred, at much larger scale, in everything that came after.
For seventeen months the sales team worked from leads generated by this pipeline. The converted prospects represented over ₪50M in new asset-management client portfolio. Cost: a workstation and one engineer.
The same pattern — extract, classify, deliver in the format the consumer already uses — recurred at larger scale: Tasko.ai and the Pipeline Observatory. The earliest version of that thinking lived here.