It's easy to think of AI as "the model". In my experience, the model is often the easiest part. The real work is everything around it.
Pipelines, not just predictions
At Megh Computing I worked on an end-to-end pipeline for Automatic Number Plate Recognition. Detecting and reading a plate is only one step; capturing frames, processing images, handling edge cases and returning results reliably is what makes it a product.
Hardware shapes the solution
I also worked on integrating an ASIC AI chip from AlphaICs with Megh's image application. Running models on specialized hardware forces you to think about performance, memory and compatibility from day one — constraints you never see in a notebook.
Trust is a requirement
At Neural-i, building image analytics for digital pathology, accuracy alone wasn't enough. The people using the system needed to understand and trust it, which shaped how we designed everything around the model.
The distance between a good model and a useful product is where most of the engineering lives.
Where I am now
Today I work with PyTorch, Hugging Face, AWS, Docker and Kubernetes, and I'm studying AI and Machine Learning at TU Darmstadt. The tools have changed, but the lesson hasn't: deploying AI well is as important as training it.