An end-to-end, reproducible pipeline that sources, normalises, enriches, scores, and generates personalised outreach for 291 qualified enterprise AI/ML leads — from raw data export to ready-to-send email + LinkedIn DM sequences.
Built for: P95.AI — an AI inference optimization platform
Target buyer: CTOs, VPs Engineering, Heads of AI at companies running LLMs in production
GitHub: github.com/aayush2724/LeadForge
Website: leadforge-876cf6-11q5x.thinkroot.app
| Metric | Value |
|---|---|
| Total active leads | 291 |
| Hot tier (score 65+) | 135 |
| Warm tier (score 40–64) | 57 |
| Cold tier (score <40) | 99 |
| Personalised emails generated | 50 |
| LinkedIn DMs generated | 50 |
| A/B variants designed | 40 (top 20 leads × 2 variants) |
| Pipeline stages | 11 (11/11 passing) |
| Sources | Apollo · LinkedIn Sales Nav · Crunchbase · GitHub · BuiltWith · Seed lists |
- Project Overview
- Problem Statement
- Pipeline Architecture
- ICP Framework
- Lead Scoring Model
- Outreach Generation
- A/B Testing Strategy
- Tech Stack
- Quick Start
- Automated Pipeline (n8n)
- Data Sourcing Guide
- Directory Structure
LeadForge is an intelligent lead qualification and personalised outreach system built for P95.AI. It combines multi-platform lead sourcing, Clay-powered enrichment, a 9-signal ICP scoring rubric, and GPT-4o outreach generation into a single reproducible Python pipeline — fully automated via n8n.
The system produces a fully enriched, scored, and outreach-ready lead database that can be re-run monthly with fresh data using one command:
python pipeline.pyOr triggered end-to-end in a single click using the n8n workflow.
P95.AI has a powerful product — an AI inference optimization layer that cuts LLM serving costs 30–45% with zero model changes. The challenge is finding and reaching the right enterprise buyers efficiently.
Three core problems:
-
No signal, no context — Raw lead lists lack tech stack data, hiring signals, and buying intent. Without enrichment, every outreach message is a guess.
-
Manual qualification doesn't scale — Human lead scoring is slow, inconsistent, and expensive. SDRs spend 70%+ of their time researching instead of selling.
-
Generic outreach fails — Non-personalised cold emails get ignored, flagged as spam, and permanently damage sender reputation. A CTO running Ray for distributed ML won't respond to a generic pitch.
LeadForge solves all three with a data-driven, signal-aware pipeline.
Raw Sources
│
├── Apollo.io export → scripts/normalize_apollo.py
├── LinkedIn Sales Nav → scripts/normalize_linkedin.py
├── Seed lists → scripts/normalize_seeds.py
└── GitHub / Crunchbase / → scripts/normalize_engineer_sources.py
BuiltWith
│
▼
scripts/compile_leads.py
(merge + dedup by domain)
│
▼
scripts/prefilter.py
(hard disqualifiers + competitor detection)
│
▼
scripts/quota_check.py
(vertical + geo distribution validation)
│
▼
Clay enrichment (manual import)
→ data/enriched_leads.csv
│
▼
scripts/enrich_3b.py
(hiring signals + funding patch)
│
▼
scripts/scoring_engine.py
(115-point ICP score per lead)
│
▼
data/scored_leads.csv
(291 active, Hot/Warm/Cold)
│
┌───────┴───────┐
▼ ▼
Phase 5 Outreach Phase 6 A/B Test
phase5_outreach.csv phase6_ab_variants.csv
(50 leads) (top 20 × 2 variants)
| Stage | Script | Output |
|---|---|---|
| 1. Normalize Apollo | normalize_apollo.py |
data/raw/apollo_normalized.csv |
| 2. Normalize LinkedIn | normalize_linkedin.py |
data/raw/linkedin_normalized.csv |
| 3. Normalize Seeds | normalize_seeds.py |
data/raw/seeds_normalized.csv |
| 4. Normalize Engineer Sources | normalize_engineer_sources.py |
data/raw/engineer_normalized.csv |
| 5. Merge + Dedupe | compile_leads.py |
data/raw_leads.csv |
| 6. Pre-filter | prefilter.py |
data/raw_leads_rejected.csv |
| 7. Quota Check | quota_check.py |
data/sourcing_qa_report.md |
| 8. Phase 3A API Enrichment | enrich_pipeline.py |
data/enriched_leads.csv |
| 9. Phase 3B Intent Enrichment | enrich_3b.py |
data/enriched_leads.csv |
| 10. Lead Scoring | scoring_engine.py |
data/scored_leads.csv |
| 11. Outreach Generation | generate_linkedin_dms.py |
data/phase5_outreach.csv |
Full definition in icp_framework.md.
Target Persona:
- Title: CTO, VP Engineering, Head of AI/ML, Director of Engineering
- Company size: 200–5,000 employees
- Funding: Series B through D, or bootstrapped with >$20M ARR
- Must-haves: LLMs in production + cloud infrastructure
- Verticals: SaaS (primary), FinTech, HealthTech, Cybersec, Logistics
- Geo: US (primary), EU/UK (secondary), India seed-only
Hard Disqualifiers:
- Competitor platforms: Baseten, Modal, Anyscale, Fireworks, Together, Replicate, RunPod, HuggingFace
- Under 50 employees
- Government or defense
- No discernible LLM workload
Tier Thresholds:
- Hot: 65+ points
- Warm: 40–64 points
- Cold: under 40 points
Each lead is scored 0–115 across 9 signals:
| Signal | Max Points | Logic |
|---|---|---|
| Contact title | 25 | CTO=25, VP Eng=20, Head/Dir=18 |
| Uses LLM in production | 20 | TRUE=20 |
| Funding stage | 20 | Series C=20, Series B=15, Series D=12 |
| Employee count | 15 | 501–2000=15, 201–500=12, 2001–5000=10 |
| Active ML hiring | 10 | Hiring ML engineers detected |
| Kubernetes in stack | 8 | TRUE=8 |
| Geo tier | 8 | US=8, EU/UK=6, India=4 |
| Ray / WandB in stack | 6 | TRUE=6 |
| GitHub AI repos | 3 | Active public AI/infra repos |
Top scoring lead: Vaibhav Nivargi — Moveworks CTO — 115/115
Personalised outreach for the top 50 Hot leads using per-lead signal context.
Cold email structure:
- Hook based on detected tech stack signal (Ray → GPU efficiency, Kubernetes → infra scale)
- Hiring signal adds urgency line
- 150–200 word target, soft CTA for 15-minute demo
LinkedIn DM structure:
- Capped at 300 characters
- Same stack-personalised hook, condensed to one punch line + ask
Sample email — Vaibhav Nivargi, Moveworks:
Subject: GPU efficiency at Moveworks scale
Hi Vaibhav,
Noticed Moveworks runs Ray for distributed ML — GPU efficiency at scale
is a real challenge. You're also hiring ML engineers, which tells me
inference demand is growing.
I'm building P95.AI, an inference optimization layer that cuts LLM
serving costs 30–45% with zero model changes.
Worth a 15-minute call to see if it fits your stack?
Top 20 leads (score 88+) receive two message variants each.
| Variant A — Pain-Led | Variant B — Social Proof | |
|---|---|---|
| Hook | Specific GPU/cost pain tied to their stack | Peer company quantified result |
| Subject example | "Your Ray cluster is probably leaving 30% GPU capacity on the table" | "How a Series C AI company cut inference costs 40% in 3 weeks" |
| Target reply rate | 12% | 10% |
| Target open rate | 45% | 50% |
| Hypothesis | High-urgency, problem-aware buyers respond better | Curious/research-mode buyers engage with social proof |
Full hypothesis documentation: ab_test_hypotheses.md
Winner declared at: +5pp reply rate over 2-week window, minimum 10 sends per variant.
| Tool | Purpose |
|---|---|
| Python 3.13 | Pipeline scripting, data processing |
| pandas | Data normalization, dedup, scoring |
| Clay.com | Lead enrichment backbone — tech stack, hiring, funding, LinkedIn signals |
| Apollo.io | Primary lead sourcing (80 leads) |
| LinkedIn Sales Navigator | High-recency contact sourcing (60 leads) |
| Crunchbase | Funding and firmographic data |
| GitHub API | Engineering signal detection — AI/infra repos |
| BuiltWith | Tech stack detection — Kubernetes, Snowflake, Ray, WandB |
| GPT-4o (OpenAI) | Personalised outreach generation |
| n8n | Full pipeline automation and orchestration |
| python-dotenv | API key management |
| tqdm + rich | Progress tracking and CLI output |
# 1. Clone the repo
git clone https://github.com/aayush2724/LeadForge
cd LeadForge
# 2. Create virtual environment
python -m venv .venv
.venv\Scripts\activate # Windows
source .venv/bin/activate # Mac/Linux
# 3. Install dependencies
pip install -r requirements.txt
# 4. Configure API keys
cp .env.template .env
# Edit .env and add your keys
# 5. Run the full pipeline
python pipeline.pyNote: All output files are pre-built and committed. You can skip directly to reviewing
data/scored_leads.csvwithout running any scripts.
| Key | Required for | Where to get |
|---|---|---|
APOLLO_API_KEY |
Apollo sourcing | app.apollo.io → Settings → API |
OPENAI_API_KEY |
GPT-4o outreach generation | platform.openai.com |
GITHUB_TOKEN |
GitHub AI repo enrichment | github.com → Settings → Tokens |
CLAY_API_KEY |
Clay enrichment | clay.com → Account → API |
The full pipeline is automated via n8n, allowing one-click end-to-end execution without touching the command line.
Step 1 — Install n8n
npm install -g n8nStep 2 — Set environment variable
On Windows (run PowerShell as Administrator):
[System.Environment]::SetEnvironmentVariable("NODE_FUNCTION_ALLOW_BUILTIN", "child_process", "User")
[System.Environment]::SetEnvironmentVariable("NODE_FUNCTION_ALLOW_EXTERNAL", "child_process", "User")On Mac/Linux:
export NODE_FUNCTION_ALLOW_BUILTIN=child_process
export NODE_FUNCTION_ALLOW_EXTERNAL=child_processStep 3 — Start n8n from the project root
cd LeadForge
n8n startStep 4 — Import the workflow
- Open
http://localhost:5678 - Go to Workflows → Import
- Select
workflows/leadforge_pipeline.json
Step 5 — Execute
- Click Execute Workflow
- All 12 pipeline nodes run in sequence automatically
Manual Trigger
↓
Compile Leads ← compile_leads.py
↓
Normalize Apollo ← normalize_apollo.py
↓
Normalize LinkedIn ← normalize_linkedin.py
↓
Normalize Seeds ← normalize_seeds.py
↓
Normalize Gaps ← normalize_gaps.py
↓
Enrich Pipeline ← enrich_pipeline.py
↓
Prefilter Competitors ← prefilter.py
↓
Score Leads ← scoring_engine.py
↓
Validate Emails ← validate_emails.py
↓
Generate LinkedIn DMs ← generate_linkedin_dms.py
↓
Sanity Check ← validate_row.py
Each node outputs { success: true, output: "..." } — visible in the n8n execution log for full auditability.
The node chain above reflects the workflow shown in the n8n editor screenshot.
n8n Editor View:
The workflow executes all 12 stages in sequence. Below is a sample execution from the n8n web UI at http://localhost:5678:
Total execution time for full pipeline: ~8–10 minutes (varies based on API rate limits and enrichment latency).
Tip: To monitor live execution, open the "Executions" tab in the n8n web UI. All logs are visible in real-time as nodes complete.
Lead sourcing from Apollo.io and LinkedIn Sales Navigator requires manual CSV export due to platform API restrictions on free/standard plans. All exported files are included in data/raw/ so the full pipeline can be re-run without re-sourcing.
Apollo.io (target: 80 leads)
- Go to apollo.io → Search → People
- Job titles: CTO, VP Engineering, Head of AI, Director of Engineering
- Headcount: 200–5,000
- Industries: Computer Software, Financial Services, Healthcare
- Keywords: machine learning, LLM, AI inference, GPU, generative AI
- Exclude: baseten.co, modal.com, anyscale.com, fireworks.ai, together.ai, replicate.com
- Export CSV → save to
data/raw/apollo_pass1_*.csv
LinkedIn Sales Navigator (target: 60 leads)
- Go to Sales Navigator → Lead Filters
- Same job titles as above
- Seniority: VP, CXO, Director
- Headcount: 200–5,000
- Activity: Posted on LinkedIn
- Geography: US, UK, Germany, France, Netherlands, India
- Save list → export → save to
data/raw/linkedin_pass*.csv
After re-sourcing:
python pipeline.py| File | Rows | Description |
|---|---|---|
data/raw_leads.csv |
297 | All normalised leads pre-filter |
data/raw_leads_rejected.csv |
6 | Disqualified leads with reasons |
data/enriched_leads.csv |
297 | Post-Clay enrichment master list |
data/scored_leads.csv |
291 | Scored + tiered, all signals |
data/scoring_report.md |
— | Score distribution + top leads |
data/sourcing_qa_report.md |
— | Vertical/geo quota validation |
data/phase5_outreach.csv |
50 | Email subject, body, LinkedIn DM |
data/phase6_ab_variants.csv |
40 | A/B variant messages for top 20 |
ab_test_hypotheses.md |
— | A/B test design + success metrics |
icp_framework.md |
— | Full ICP definition + scoring rubric |
LeadForge/
├── README.md
├── icp_framework.md
├── ab_test_hypotheses.md
├── pipeline.py
├── fix_violations.py
├── requirements.txt
├── .env.template
│
├── workflows/
│ └── leadforge_pipeline.json ← n8n workflow (import to re-run)
│
├── data/
│ ├── SCHEMA.md
│ ├── raw_leads.csv
│ ├── raw_leads_rejected.csv
│ ├── enriched_leads.csv
│ ├── enriched_leads_3a_backup.csv
│ ├── scored_leads.csv
│ ├── scored_leads_validated.csv
│ ├── phase5_outreach.csv
│ ├── phase6_ab_variants.csv
│ ├── scoring_report.md
│ ├── sourcing_qa_report.md
│ ├── enrichment_run_log.md
│ ├── enrichment_3b_log.md
│ └── raw/
│ ├── apollo_normalized.csv
│ ├── apollo_pass1_us_saas_fintech.csv
│ ├── apollo_pass2_eu_uk.csv
│ ├── apollo_pass3_healthtech_cybersec.csv
│ ├── apollo_pass4_flex.csv
│ ├── linkedin_normalized.csv
│ ├── linkedin_pass1_us.csv
│ ├── linkedin_pass2_eu.csv
│ ├── linkedin_pass3_india.csv
│ ├── seeds_normalized.csv
│ ├── seeds_raw.csv
│ ├── engineer_normalized.csv
│ ├── aayush_normalized.csv
│ ├── builtwith_raw.csv
│ ├── crunchbase_raw.csv
│ ├── github_raw.csv
│ ├── cyber_normalized_gap.csv
│ ├── cyber_people_gap.csv
│ ├── ecommerce_normalized_gap.csv
│ ├── ecommerce_people_gap.csv
│ ├── healthtech_normalized_gap.csv
│ ├── healthtech_people_gap.csv
│ ├── logistics_normalized_gap.csv
│ └── logistics_people_gap.csv
│
├── scripts/
│ ├── compile_leads.py
│ ├── normalize_apollo.py
│ ├── normalize_linkedin.py
│ ├── normalize_seeds.py
│ ├── normalize_engineer_sources.py
│ ├── normalize_gaps.py
│ ├── prefilter.py
│ ├── quota_check.py
│ ├── enrich_pipeline.py
│ ├── enrich_3b.py
│ ├── scoring_engine.py
│ ├── generate_linkedin_dms.py
│ ├── validate_emails.py
│ ├── validate_row.py
│ └── enrichers/
│ ├── __init__.py
│ ├── apollo_enricher.py
│ ├── crunchbase_enricher.py
│ ├── github_enricher.py
│ └── jobs_enricher.py
│
├── docs/
│ ├── team_roles.md
│ ├── clay_setup.md
│ ├── n8n_workflow_execution.png
│ └── clay_screenshots/
│
└── logs/