/Jace RoldanA minimal, responsive and feature-rich Jekyll theme for technical writing. 2025-08-07T15:22:20+00:00 Jace Roldan / Jekyll © 2025 Jace Roldan /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png Sari Sandbox: A Virtual Retail Store Environment for Embodied AI Agents2025-08-07T00:00:00+00:00 2025-08-07T15:21:55+00:00 /posts/sari-sandbox-designing-a-virtual-retail-environment/ Jace Roldan We published a new paper on the design of a virtual retail store environment for embodied AI! Link to our website: https://sarisandbox.github.io/. CuatroLLM niche translations with adaptive in-context learning: exploring English-Filipino2025-01-06T00:00:00+00:00 2025-07-16T02:59:27+00:00 /posts/cuatro-llm-fine-tuning/ Jace Roldan CuatroLLM is a 1.3B parameter four-language machine translator pre-trained on a 300B-token dataset, TransWeb-Edu, which the authors posted on arXiV last October 2024. Summary This mini-project focused on CuatroLLM, a 1.3B parameter model pre-trained on four languages: English, French, Spanish, and German. We replicated its baseline results on complex reasoning tasks. Fine-tuning CuatroLLM wit... A YOLO Segmentation Model for Grocery Items!2024-11-30T00:00:00+00:00 2025-07-03T06:10:06+00:00 /posts/writing-a-YOLO-object-detection-grocery-item/ Jace Roldan Summary An experiment to train a model to detect Filipino grocery items using YOLO. YOLO v11 model (X-size) is trained in this experiment for object detection and segmentation tasks for common Filipino grocery items under a variety of model configurations and data augmentation techniques. An inference application is deployed via tunneling on a Gradio interface. Prompt The idea was simple — c...