Generative AI for Interior design: 2nd place

I participated in the competition related to filling the empty room. During the competition, I learned how to control the Stable Diffusion models, how to gather training data, and how to train it properly. The most challenging issue was room layout preservation. More info at the blog post.

2nd place in Visual Search competition

The aim was to develop an image search engine for e-commerce. The input is a user image and we should return the most similar items from the store. Everything should work in nearly real-time. Apart from learning an embedding model, this task requires handling a huge scale of data as we had > 1B imgs

OCR for receipts

The task was to create a bespoke solution for Indonesian receipts. Open-sourced libraries weren't working well there. I was evaluating a couple of open-source solutions (Tesseract, Ocropy) and implementing my own OCR engine.

Satellite image segmentation

The idea was to find wetlands in the USA using satellite images from the Sentinel-2 dataset and open-source labels. The majority of time was taken by data preparation steps, to make the whole system scalable, easy to maintain, and provide enough accuracy.