Case studies / Proof of concept (PoC) with generative AI
The challenge
Ads were running but not reaching the right people
The client advertised products and services on several websites, but its ad strategy wasn't getting the results it wanted. Less relevant ads reached the wrong audiences. Click-through and purchase rates both fell. The client needed a system that could identify user segments accurately and show each one the right ads.
What we did
Group users by behavior, then choose their ads
- Visit data collection and analysis: The system collects and analyzes website visitor data, using metadata, purchase history, IP addresses, cache data, and more.
- Grouping users with generative AI: Generative AI sorts users into groups based on the analyzed data.
- Showing the best ad: Each group sees the most relevant ad in the ad slot.
Results after launch
Click-through rate up about 30%
- Higher click-through rate: With the new way of grouping users and showing ads, the click-through rate rose about 30% on average from the previous month.
- Higher purchase rate: Showing the most relevant ads raised the purchase rate by 15%, with the clearest gains in target customer groups.
- Better engagement rate: Users became more likely to take an interest in ads and interact with them, and the engagement rate rose 25% on average.
What we did next
An AI chatbot for internal documents and content
Many internal documents were shared and changed often, which made changes hard to track. Writing articles and other content for the website also took a lot of time. We added automated answers to the existing chatbot, so staff find information faster and create content more efficiently.
Engagement: up to 2 people × 4 months
- Management module: Handles document, collection, and user data; relays requests to the large language model (LLM) and orchestrator; and manages authentication and permissions.
- Orchestrator: Loads documents from data sources and coordinates the processing and data flow when a question comes in.
- Task queue: Queues requests to the LLM engine and notifications to the management module, and monitors the queue with Flower.
- Text search: Turns text into vectors with embeddings and stores the vectors and metadata for search.
- Easier change tracking: Efficient change tracking means fewer mix-ups and less duplicated work.
- Article and content creation time cut by 30%: The latest information is shared quickly, and the time saved goes to other work.
System overview
From visit data to targeted ads
← Scroll sideways to see more →
Technologies used
The stack for phase 1 and phase 2
Phase 1: Personalized ad prediction system
| Languages | Python |
|---|---|
| API | FastAPI |
| Packaging | Docker, Ansible |
| Deep learning frameworks | TensorFlow, PyTorch |
| MLOps | Kubeflow (Kubernetes), Seldon Core / Load balancing: Nginx / Message queues: Kafka, Redis, RabbitMQ |
Phase 2: General-purpose AI chatbot
| Languages | Python, LangChain |
|---|---|
| Frontend | React (Ant Design) |
| Management | Loopback |
| Database | MongoDB |
| Processing tools and services | Celery worker, Flower, Llama 2, FastAPI |
| Natural language processing | Qdrant (vector database, text embeddings) |
