LeadFlow NL | Prince Festus

Overview

What LeadFlow does.

In 5 days, I built a system that finds Dutch cleaning companies, researches each one individually, and sends their owner a cold email that reads like it was written by someone who actually looked at their business. It follows up, tracks replies, and reports daily. Research, drafting, scheduling, and follow-up are automated. Human involvement is limited to final approval before emails are sent.

The emails follow strict rules: maximum 6 lines, one genuine observation, one question, no pitch, no mention of services. The goal is to start a real conversation, not close a sale in the first message.

The problem

Why generic outreach fails.

Most cold outreach is either manual and slow or automated and generic. LeadFlow was built to automate personalisation without sacrificing quality.

Why I built it

From manual to automated.

I was manually researching companies, writing outreach emails, and tracking replies. The process worked but didn't scale. LeadFlow was built to automate research, drafting, and follow-up while keeping a human approval step for quality control.

Outcomes

Early results.

9

Emails sent in testing phase

2

Positive replies from business owners

22%

Initial reply rate — 2 replies from 9 test emails

2 replies from 9 emails sent in initial testing (22% reply rate). Sample size is intentionally small. Larger-scale testing is ongoing. These results indicate early signal, not statistically proven performance.

In action: two real leads

Real emails. Real replies.

01JP Cleaningservice

LeadFlow researched JP Cleaningservice's specialisation in facade cleaning and certified glasswork, then drafted this email. Specific, short, no pitch.

They replied within hours, sharing their story and inviting a personal conversation. The system hadn't pitched anything. Just asked a genuine question.

02Lotos Clean

LeadFlow spotted Lotos Clean's niche in hotel cleaning in Amsterdam — unusual for a cleaning company. That became the entire angle of the email.

They asked what the interest was. The follow-up explained the context. No hard sell, just a conversation about their growth.

Their final reply: "I'm always looking for new clients. If you can offer something interesting that works, I'm willing to explore." Qualified lead. Fully automated.

Human in the loop

Automation scales. Humans judge.

Automation handles scale. Humans handle judgment.

Every email goes through a Telegram approval gate before it's sent. The operator sees the draft, can approve it as-is, edit the subject line, rewrite the body, or reject it entirely. Nothing leaves without a human sign-off.

The bot queues drafts for review. The operator can approve, edit subject line, edit message body, or reject. All from inside Telegram.

@leadflownl_bot, the dedicated approval bot. Every outreach passes through here before it reaches a real inbox.

System Architecture

How lead discovery, research, AI generation, approval, and delivery work together.

Google Maps

GoudenGids

Bedrijvengids

Dutch Business Directories

↓

Lead Discovery Engine

Python · Scraper

↓

PostgreSQL Lead Database

Deduplication · City Tracking

↓

Qualification Engine

Hot (80+) · Warm (40+) · Discard

↓

Research Pipeline 6 parallel tasks

Website Analysis

Reviews

Social Profiles

Job Posts

News Mentions

Prospeo Enrichment

↓

Claude

Draft Generation

↓

GPT-4o-mini

Quality Review

↓

Telegram Approval Bot

Human sign-off required before send

↓

Email Delivery Layer

Gmail + Outlook Rotation · Daily Caps

↓

Reply Tracking

→

Daily Reporting

How it works

Six stages, fully automated.

01

Discovery

Scrapes GoudenGids, Bedrijvengids, and Google Maps across all 342 Dutch municipalities in population order. Captures name, address, phone, website, and email. Deduplicates automatically and tracks which cities are exhausted.

PythonPostgreSQLGoogle MapsGoudenGidsBedrijvengids

02

Qualification

Each company is scored on years in business, employee count, and specialisation. Hot (80+) and warm (40+) move forward. Everything below is filtered out. The pipeline only works on companies worth contacting.

Pythonscore_company.py

03

Research

Six parallel tasks run against each qualified company: full website analysis, Google review scraping, social media presence, job postings, news mentions, and contact enrichment via Prospeo to find a founder's personal email rather than a generic inbox.

PythonProspeo API

04

Drafting

Claude writes the personalised Dutch cold email from the research profile. GPT-4o-mini then acts as an independent editor, catching generic phrasing, tone drift, and anything that sounds like it was written by a machine. Two models because one model cannot honestly judge its own output.

Claude APIGPT-4o-mini

05

Approval

Draft surfaces in Telegram. Operator approves, edits, or rejects. Nothing sends without sign-off.

Node.jsTelegram Bot API

06

Sending & Follow-up

Emails rotate across Gmail and Outlook accounts with enforced daily caps (5 → 15 → 30/day per phase). Non-replies trigger automated follow-up sequences. A master cron scheduler runs everything: scraping at 7AM, reporting at 6PM. Fully unattended on Railway.

Node.jsGmail APIRailway

Engineering challenges

What was actually hard to build.

01

Multi-source scraping without duplicates

Synchronising company data from three sources (Google Maps, GoudenGids, and Bedrijvengids) required building a deduplication layer that could match companies with slightly different names, addresses, or phone numbers across sources without losing legitimate entries or creating duplicates in the database.

02

Email deliverability under rotation

Managing Gmail and Outlook account rotation while respecting per-account daily send limits, avoiding sender reputation damage, and handling OAuth token refresh across multiple accounts required careful orchestration, especially when Railway's outbound SMTP was blocked and OAuth2 had to replace it entirely.

03

AI quality control at the output level

Using separate generation and review models (Claude to write, GPT-4o-mini to edit) reduced generic outputs and helped maintain consistent Dutch tone. A single model reviewing its own output consistently missed tone drift and marketing language it had introduced. The two-model approach caught what one model alone would not.

04

Fault tolerance across an unattended pipeline

Handling API timeouts, exhausted search regions, Prospeo returning no results, and incomplete company data, all without the pipeline silently continuing or crashing. Every failure mode required a defined degradation path: skip and log, alert and pause, or retry with backoff. A system running unattended at 3AM needs explicit failure behaviour at every stage.

Key decisions

The choices that define the product.

01

Honesty over volume

The system was built with one hard constraint: never send a fake compliment, a made-up statistic, or a generic email dressed up as personal. One dishonest email doesn't just get ignored. It poisons every email that follows. Every research claim is verified before it becomes a sentence in an outgoing message.

02

Two models, not one

Claude writes. GPT-4o-mini edits. The reason: a model cannot honestly evaluate its own output. It has too much attachment to what it generated. A second model with no context on the original draft will catch what the first one misses: generic phrasing, marketing drift, unnatural Dutch tone. This is the core of why the emails read like a human wrote them.

03

Loud failures only

A system running unattended at 3AM with no one watching has two acceptable states: working correctly, or alerting loudly. Scraping fails, APIs time out, cities exhaust, Prospeo finds nothing. Every failure mode was designed to degrade gracefully or surface immediately. Silent failures are the most dangerous kind.

Built with

The stack behind it.

AI Models

Claude API GPT-4o-mini

Infrastructure

Node.js Python PostgreSQL Railway

Integrations

Gmail API (OAuth2) Telegram Bot API Prospeo API

What this tells you

What I bring to a team.

LeadFlow isn't just a technical project. It's proof that I can identify a real business problem, design a solution from scratch, build and deploy it in production, and iterate until it works. I built this in 5 days, working alone. The system automates lead discovery, research, drafting, approval workflows, and follow-up scheduling while providing monitoring and failure alerts.

If you're looking for someone who ships, thinks in systems, and doesn't stop at a working prototype. That's what I do.