← All projects
PUZZLE CRM
SAN FRANCISCO, CA
BUILT 2025–2026
AI Automation Portfolio — Case Study 01

I built a 5-channel
outreach system

Built for a health & wellness practice in SF. Two systems: a referral pipeline across 5 channels and a reusable KOL outreach playbook with AI-scored influencer matching. Templatized for any brand. Fully automated.

Lead Discovery Lead Scoring Email Automation SMS Outreach Instagram DM Web Form Fill Reply Detection KOL Playbook AI Brand-Fit Scoring Life Signal Detection CRM Database Campaign Tracking Lead Discovery Lead Scoring Email Automation SMS Outreach Instagram DM Web Form Fill Reply Detection KOL Playbook AI Brand-Fit Scoring Life Signal Detection CRM Database Campaign Tracking

Most people talk about using AI. I shipped a system that actually runs.

Outreach Channels
📧
Email Outreach
Gmail IMAP + SMTP

AI-personalized cold emails with daily reply detection. Auto-classifies responses as interested, not interested, or neutral. Updates CRM score in real time.

💬
SMS / iMessage
Twilio + iMessage

Text outreach for event mixer campaigns. Detects STOP / unsubscribe replies and permanently flags contacts so they're never messaged again.

📸
Instagram DMs
Playwright CDP — 26 sent

Rate-limited to 10/day. Pre-engagement (like + comment) before DM. Human-like delays. Anti-ban compliance built in.

🌐
Web Form Fill
Playwright automation — 103 targets

Browser automation fills contact forms on clinic websites with 90-second+ human-like delays between submissions. Screenshots saved for verification. Fallback to email if form fails.

📍
Lead Discovery
Google Places + NPPES + Instagram

Three discovery pipelines: neighborhood search across 8 professions × 6 SF areas, NPPES provider registry, and Instagram social graph crawl from @puzzletcm.

The Pipeline

Every contact flows through five automated stages before a human ever sees a reply.

01
Discover
  • Google Places API
  • NPPES registry
  • Instagram graph
  • Apify scraping
02
Enrich
  • Website email extraction
  • Fuzzy deduplication
  • Credential detection
  • Distance calculation
03
Score
  • Tier 1: warm (social)
  • Tier 2: nearby (geo)
  • Tier 3: cold (broad)
  • Dynamic scoring
04
Outreach
  • Email + SMS + DM
  • Web form fill
  • Rate limiting
  • Personalization
05
Track
  • Reply detection
  • Score updates
  • Unsubscribe guard
  • Activity log

Built once.
Runs for any brand.

A 6-phase influencer outreach methodology. Templatized campaigns, AI-scored brand fit, automated daily sends. Designed as a reusable playbook, not a one-off script.

01
Define
Target audience, competitor accounts, campaign angle. Each campaign targets a specific persona with a specific reason to care.
02
Collect
Modash API pulls collaborator data from competitor accounts. Zero IG visits needed at this stage.
03
Score
AI brand-fit scoring on a 1-5 scale. Life signal detection from bios. Human review on borderline 3s only.
04
Template
DM + email templates. Show 2-3 examples for approval, then batch generate all. Never review every message individually.
05
Send
Rate-limited daily sends via LaunchAgent automation. One IG visit per profile. Bio capture + send in a single pass.
06
Track
CRM scoring: replies, appointments, conversions, declines. Score system prevents re-contacting cold or declined leads.
How it thinks
Life Signals
Detects parents, athletes, couples, event organizers from bio text. Adjusts the pitch angle automatically per signal.
Brand Fit
AI scores each influencer 1-5 on lifestyle match. Score 3 gets flagged for human review. Scores 1-2 auto-cut.
Platform Safety
All data collection via Modash APIs. IG is only touched for sending. Rate-limited, anti-ban compliant, screenshot-verified.
By the numbers
0
NPPES leads
in discovery pipeline
0
playbook phases
per KOL campaign
0
KOL campaigns
designed and shipped
0
influencer profiles
pulled from Modash
0
search combinations
8 professions × 6 areas
0
outreach channels
in one system

Tech
Stack

Python for data pipelines. Node.js for browser automation. SQLite keeps it all honest. No fancy infrastructure.

Core Language
Python 3 Node.js ESM SQLite 3
Browser Automation
Playwright Chrome CDP Apify
APIs + Data
Google Places API Google Maps NPPES Registry Modash Gmail IMAP Twilio
Discovery + Scraping
instagrapi BeautifulSoup4 httpx geopy thefuzz
Compliance + Safety
Rate limiting Unsubscribe detection Dry-run mode Screenshot verify
Testing
pytest pytest-mock
Built by

Jenny
Chen

GTM operator turned builder. I've taken companies from 0 to $5M ARR. Now I build the systems I always wished existed. This is one of them. Currently targeting GTM + Partnerships roles at Series A-B AI companies.

Puzzle CRM
v1.3.4
San Francisco
2025–2026