AVN Detection System

The Challenge

Amazon’s vehicle inspection process required manual review of hundreds of thousands of vehicle images annually. Inspectors had to identify whether vehicles displayed required AVN (Amazon Vehicle Network) identification markers—a tedious, time-consuming task that created bottlenecks and significant operational costs.

The inspection team was drowning. Accuracy was inconsistent. Costs were scaling linearly with fleet growth.

The Constraint

Traditional approaches wouldn’t work:

• Engineering backlog: The ML team had an 18-month roadmap. This wasn’t on it.
• Budget: No approved budget for a dedicated engineering solution.
• Time: The problem needed solving in months, not years.
• Resources: I had no engineering team assigned to this problem.

The conventional wisdom said: wait for engineering, or accept the manual process.

The Approach

I built it myself. Using AI tools as both teacher and collaborator:

Week 1-2: Problem Definition

• Analyzed existing inspection workflow and failure modes
• Identified the specific detection requirements (what constitutes a valid AVN marker)
• Determined accuracy thresholds needed for auto-approval

Week 3-8: System Development

• Used Claude and GPT-4 to learn computer vision fundamentals
• Built detection pipeline using pre-trained models + fine-tuning
• Developed validation framework to measure accuracy
• Iterated through multiple architectures until hitting accuracy targets

Week 9-12: Production Integration

• Integrated with existing inspection workflow
• Built monitoring and alerting for edge cases
• Deployed to production with gradual rollout
• Validated performance against manual inspection baseline

Week 13-16: Optimization & Handoff

• Optimized inference costs
• Documented system for maintenance
• Trained operations team on monitoring

The Outcome

Primary Metrics:

• 333x cost reduction compared to manual inspection process
• 100% accuracy on auto-approved detections (conservative thresholds)
• 4 months from concept to production deployment

Operational Impact:

• Eliminated inspection bottleneck for AVN compliance
• Freed inspection team to focus on complex edge cases
• System handles growing fleet volume without linear cost increase

What Didn’t Work:

• First two model architectures failed to generalize across lighting conditions
• Initial accuracy thresholds were too aggressive—had to calibrate for production reliability
• Underestimated edge cases involving partial occlusions

What This Proves

This wasn’t about being a coding genius. I’m not one.

This was about:

1. Domain expertise: Understanding the actual problem deeply enough to specify a solution
2. AI tools as teachers: Learning computer vision concepts through iterative dialogue with AI
3. Discipline: Testing, validating, iterating until it worked—not shipping something half-baked
4. Production mindset: Building for reliability, not demo-ware

The 333x cost reduction wasn’t achieved by a team of ML engineers. It was achieved by a PM who refused to accept “wait for engineering” as an answer.

---

Note: Specific implementation details and proprietary methodologies omitted per confidentiality requirements. Metrics represent approximate impact.