Ad-Ops Autopilot: AI Ad Copy Generation Engine
Autonomous ad copy generation system for Facebook and Instagram that generates, evaluates, and iteratively improves ad copy using Pareto-optimal selection and a quality ratchet that ensures standards only go up over time.
Ad-Ops Autopilot: AI Ad Copy Generation Engine
Role: AI Engineer
Program: Gauntlet AI — 2-Month Immersive for AI Engineers
Live Demo: adautomationengine.vercel.app
GitHub: github.com/alediez2048/nerdy
Tools: Python, Google Gemini API, Chart.js, pytest, YAML, JSONL
Overview
Ad-Ops Autopilot is an autonomous ad copy generation system built for Facebook and Instagram campaigns. It takes a creative brief and runs it through a fully automated pipeline: Brief → Expand → Generate → Evaluate → Publish or Regenerate.
The system generates multiple ad copy variants, scores them across five quality dimensions, and uses Pareto-optimal selection to pick the best variant — with a quality ratchet that ensures standards only go up over time.
How It Works
- Brief Intake: Takes a creative brief with target audience, brand voice, and campaign goals
- Expansion: Enriches the brief with market context and competitive intelligence
- Generation: Produces 3-5 ad copy variants per cycle
- Evaluation: Scores each variant across 5 dimensions — Clarity, Value Proposition, CTA, Brand Voice, and Emotional Resonance
- Selection: Pareto-optimal selection picks the dominant variant with no dimension regression
- Iteration: If quality thresholds aren't met, the system regenerates with targeted improvements
Key Features
- Evaluator-First Architecture: Scoring system was built and calibrated to 89.5% accuracy before the generator
- Pareto-Optimal Selection: Multi-dimensional quality optimization, not single-score ranking
- Quality Ratchet: Standards only increase over time — no quality regression allowed
- 8-Panel HTML Dashboard: Real-time visualization of KPIs, iteration cycles, quality trends, token economics, and system health
- Full Audit Trail: Append-only JSONL ledger for checkpoint-resume and forensic replay
- 670+ Tests: Golden set regression, adversarial boundary, and pipeline integration coverage
- SPC Monitoring: Statistical Process Control for detecting system health anomalies
Tech Stack
| Category | Details |
|---|---|
| Language | Python 3.10-3.12 |
| AI/LLM | Google Gemini API (model routing: Pro for improvable-range scores) |
| Dashboard | Single-file HTML with embedded CSS/JS + Chart.js |
| Testing | pytest (670 tests) |
| Data | YAML config, JSONL ledger/logging |
| Deployment | Vercel |
