2025-09-25 –, Apollo
In this talk, we'll share how we're using computer vision to automate stock counting, right on the conveyor belt. We'll discuss the challenges we've faced with the hardware, software, and GenAI components, and we'll also review our own benchmark results for the various state-of-the-art models. Finally, we'll cover the practical aspects of GenAI deployment, including prompt optimization, preventing LLM "yapping," and creating a robust feedback loop for continuous improvement.
At Picnic, we deliver groceries to customers across three countries. To ensure we deliver what we promise, we rely on precise stock tracking. But with millions of items moving through our highly automated fulfillment center in Utrecht every day, manually counting stock is incredibly labor-intensive.
In this talk, we'll share how we're using computer vision to automate stock counting, right on the conveyor belt. From capturing high-quality images of moving products to evaluating and deploying state-of-the-art models, this project challenged us across the stack: hardware, software, and GenAI.
We'll explore topics such as:
* Why YOLO didn't work for us
* Benchmarking the latest (GenAI) vision models
* Optimizing prompts with DSPy
* Preventing unnecessary LLM “yapping”
* Capturing high-quality images of items in motion
This session blends practical engineering insights with cutting-edge AI tools making it interesting for anyone applying computer vision to real-world problems.
Sven Arends is a Senior Machine Learning Engineer at Picnic, where he is currently developing novel LLM applications. With a strong background in machine learning and software engineering, Sven has contributed to a wide range of projects: from predicting delivery times to forecasting buying behaviour and implementing MLOps tooling within Kubernetes. He holds a Master’s degree in Computer Science.
In his free time, Sven enjoys creating small embedded devices or visiting the sauna: a tradition he adopted during his time in Finland.