At this startup focused on efficient, edge-deployed AI models (it was founded by the inventor of Extreme Learning Machines (ELMs)), I advanced from a junior engineer to leading two teams by intitiating a few key projects across our object detection pipeline.

My primary work centered around optimizing the performance of lightweight object detection models, which I did through a series of improvements: SimCLR-inspired data augmentations, loss-dependent learning rate scheduling, and a custom SGD implementation to facilitate hyperparameter tuning. These, along with improvements to our datasets, increased our model performance on YOLOX tiny from 0.5 to 0.85 mAP across multiple datasets. I also worked on deploying these models in a quantized RKNN format (for inference on Rockchip NPUs).

The most impactful contribution I made was probably to the dataset curation process. In order to quantify the biasness of our datasets, I developed a latent space analysis and clustering, proving that the dataset had too many near-duplicates and allowing their pruning to be automated. Then, to make up for this loss of data and beef up our datasets, I created a semi-supervised pipeline for dataset annotation, also replacing the manual labelling exercises (which I dreaded) with a self-training system that iteratively labeled and trained a large vision model using scraped data. Combined, these tools allowed us to rapidly curate datasets that were 3-5x larger than our original ones (and with greater diversity too!). For this, I was given two different teams to scale these tools up.

These contributions didn't just improve our R&D metrics, but also laid the groundwork for productizing our object detection solutions.