The Business Challenge
The business situation the organisation faced before the project began.
Verifying that advertisements appear correctly and as agreed is traditionally a manual task: someone has to look at images and confirm what is shown. At volume this is slow, inconsistent and hard to scale, and generic image tools struggle with the variety of real-world ad content.
Why AI Was the Right Solution
Why AI was the right approach — and what alternatives were considered.
Ad creatives vary widely in layout, text and imagery, so fixed rules or simple image matching miss too much. A vision-language model can understand what an image shows and how it relates to what was expected, which is what verification requires.
How the Solution Was Delivered
Discovery, design, development, testing and rollout — the journey, not the tooling.
- 1Mapped the manual verification process and what reviewers needed to confirm in each image.
- 2Selected the Qwen vision-language model for its ability to understand varied visual content.
- 3Set up GPU inference on two Tesla V100 32GB GPUs.
- 4Built pipelines for image recognition and ad verification.
- 5Validated outputs against expected results and refined the pipeline.
Key Technical & Architecture Decisions
Architecture, model selection, workflow and trade-offs.
Vision-language model rather than a narrow classifier, to handle varied ad content.
Qwen as the model, run on dedicated GPUs for control over performance and data.
Two Tesla V100 32GB GPUs to fit the model and keep throughput high.
Pipeline design that separates recognition from verification logic.
Challenges & How They Were Solved
Obstacles hit along the way and how they were overcome.
GPU memory limits for a large vision-language model: handled by sizing the deployment across two 32GB GPUs.
Variety in real ad content: addressed by using a model that understands context instead of fixed templates.
Consistency of results: improved by validating outputs and refining the pipeline.
User Adoption & Change Management
How users responded — training, change management and feedback loops.
Reviewers moved from checking every image to handling the cases the pipeline flags, which makes their time more valuable and builds confidence in the automated results.
Business Outcomes & Impact
The measurable outcomes from the underlying AI Deployment — with the story behind them.
Manual ad verification replaced by an automated vision-language pipeline running on 2x Tesla V100 32GB GPUs.
The story behind the numbers
Ad verification moved from manual inspection to an automated pipeline. Checks are faster and more consistent, and capacity grows with GPU resources rather than reviewer time, freeing people to focus on edge cases instead of routine checks.
Lessons Learned
What surprised the team, what worked well, and what would be done differently.
- Vision-language models handle varied visual content better than fixed rules.
- GPU capacity planning is a first-class design decision.
- Human review of edge cases keeps automated verification trustworthy.
Future Opportunities
Where this solution could go next.
Confidence scores to route uncertain cases to human review.
Batch scheduling to maximise GPU utilisation.
Support for video ads.
Dashboards showing verification volume and outcomes.
Professional Reflections
Guidance for another builder tackling a similar problem.
Choose models for the variety of real-world data, not the clean examples. Plan GPU capacity early for vision-language models, and keep humans in the loop for low-confidence cases while trust in automation builds.