Home/Work/AI-Powered Data Pipeline
● Case Study 01 — Lead Generation

AI-Powered Data Pipeline

A fully autonomous, 4-stage system that turns thousands of raw web records into a handful of high-value leads in your inbox — every single day, with no manual review.

Client
B2B agency (under NDA)
Role
Sole engineer · end-to-end
Timeline
3 weeks to production
Status
Live · running daily
The challenge

Hours of manual prospecting, every morning.

The client's team started each day the same way: manually combing job boards, news feeds and directories for companies that might need their service — copying details into a spreadsheet, guessing at relevance. Slow, inconsistent, and impossible to scale.

Before
  • ~6 hours of manual research spread across the team, daily
  • Inconsistent criteria — every person judged "good lead" differently
  • Best prospects buried under hundreds of irrelevant results
  • No way to scale without hiring more people
After
  • Zero manual research — the pipeline runs unattended every morning
  • One consistent, tunable scoring standard applied to every record
  • Top 3 leads delivered straight to Telegram, ranked and explained
  • Scales to new sources by changing a config — not headcount
The solution

Four stages, each doing the cheapest possible work first.

The key insight: never pay an API to look at noise. Each stage aggressively narrows the funnel, so expensive AI reasoning only ever touches records that already passed cheaper local checks. The result is near-zero running cost.

01
Playwright

Scrape

Headless Chrome crawls 15+ sources daily, extracting structured data from JS-heavy sites.

02
Ollama · local

Filter

A local Mistral model pre-screens every record, eliminating 80%+ of noise for free.

03
Claude API

Evaluate

Claude scores the survivors against custom criteria and extracts the key signals.

04
Telegram

Notify

The top-ranked leads arrive as a formatted message — company, score, direct link.

See it run

The funnel, live.

This is the real shape of a daily run. Watch thousands of raw records collapse into three leads worth a human's attention. Hit Run pipeline.

~/lead-engine — autonomous
Scrape
Playwright
0
records collected
Filter
Ollama
0
records survived
Evaluate
Claude
0
records scored
Notify
Telegram
0
leads delivered
Idle — ready to run
The impact

What changed for the client.

~6h→0
Daily manual research, eliminated
4,280
Records processed automatically per day
$0.04
Average API cost per full daily run
24/7
Runs unattended — no human in the loop
A closer look

Inside the system.

Drop in your own screenshots — dashboard, Telegram delivery, scoring detail.

"We went from dreading the morning lead hunt to just checking Telegram over coffee. The quality is better than what we found by hand — and it never takes a day off."
— Head of Growth, B2B agency (name withheld under NDA)
Built with

Tech stack.

Claude APIOllama (Mistral)Node.jsTypeScriptPlaywrightMongoDBTelegram Bot APIDockerCron / scheduling

Have a workflow that eats your team's mornings?

If a task is repetitive, rule-based and data-heavy, it can probably run itself. Tell me about it — I'll reply with an honest assessment within 2 hours.

Start a project →