AITownCrier combines a fine-tuned stable diffusion model with OpenAI's chatGPT to generate and curate unique visuals and narratives from scraped news stories worldwide. With the integrated system set up on AWS, these reimagined news stories are automatically posted to Twitter.
Adopting the approach outlined in The Devil is in the Details paper, I finetuned a Segformer model to deliver zero-shot segmentation. By converting image labels to sentence embeddings, the model could identify various architectural components across diverse building styles without prior exposure to these elements.
Miniaturisation project for object detection models (YOLOv5) aimed at edge compute devices. With the aid of TensorRT, I re-implemented and miniaturised models to be deployable on Jetson devices, and carried out evaluations of the models' performance under these new constraints.
Managed and developed an end-to-end Machine Learning pipeline, built on Kubernetes, that addressed all facets of Machine Learning workflows via RESTful APIs. This included data ingestion, transformations, model training, evaluation, and deployment. Provided a comprehensive, scalable, and efficient way to handle ML tasks within the organisation.
Leveraging Blender and open-source toolkits, I developed a swagger documented microservice for generating synthetic images to curate bespoke training sets. This approach helped to augment scarce and unique data sets, enhancing the efficacy of object detection algorithms.
Implemented data fusion methodologies for training object detection networks.