01The problem
The master sheet had duplicate entries, and follower counts had to be looked up by hand, one handle at a time. There was no written reason why one influencer ranked above another, so every shortlist conversation started from opinion.
The campaign needed something it could spend against with confidence: fewer names, better names, and a reason attached to each one.
02The funnel
03How it works
- Clean and dedupe. A Python pipeline normalised the 1,463 records across 28 states and removed the duplicates.
- Automate the research. Follower counts for 1,400+ handles were scraped instead of looked up. That removed more than 90% of the manual research.
- Score with reasons. A rule-based classifier wrote a rationale file for each candidate it scored, 645 in all, saying which rules applied and why.
- Review in batches. The 645 went to subject experts in 15 batches.
- Filter to the shortlist. Contact, follower and geography criteria across 27 fields cut the list to 273, tiered and searchable by state.
04Decisions
D1Rules instead of a black-box score
A rule a reviewer can read beats a number they have to trust. Every rationale names the rules that fired, so disagreement is about a specific rule, not a gut feeling.
Trade-offRules miss nuance a person would catch. That's what the expert review was for.
D2Fifteen small reviews, not one big one
About 43 records per batch, small enough for an expert to finish in one sitting.
Trade-offSlower end to end than one pass, but mistakes surfaced early instead of all at once.
D3Never overwrite the master
I kept the master sheet's versions across 10 backup states. Any cut could be traced back or undone, and nothing was lost.
05Side quest: 50 reels
Picking the right people is half the job. The other half is knowing what works for them. I built the tooling for a content audit of 50 Instagram reels: scrape them, transcribe them in bulk with the Gemini Flash batch API, and look for patterns in how top creators open their videos. The output was a list of hooks the campaign could brief creators with.
06Results
07Under the hood
What I designed or built myself:
- clean A Python pipeline that normalises and dedupes 1,463 records across 28 states and 27 fields.
- scrape Automated follower counts for 1,400+ social handles.
- classify A rule-based classifier that writes one rationale JSON per candidate, 645 in all.
- review The 645 split into 15 batches for subject-expert validation.
- audit 50 reels scraped and transcribed with the Gemini Flash batch API.
- safety The master sheet versioned across 10 backup states.
08Looking back
I'd put the rules in front of the reviewers before writing any code. Agreeing on them first costs one meeting. Changing them after 645 rationales exist costs a rerun.