About me

An AI-driven application that automates the end-to-end generation of compliance report PowerPoint slides from custom data inputs.

The solution applies robust software engineering practices and large language models (LLMs) to analyze user-provided data and transform it into structured, client-compliant presentations. In addition, a human-in-the-loop (HITL) workflow was implemented, enabling compliance managers to refine and edit the generated PPTX through prompt-based requests. By replacing a fully manual reporting process, the application significantly improves productivity, reducing report creation time and operational overhead.

A video captioning system that generates detailed captions for military scenarios.

The system should be able to generate detailed captions that describe the actions and events taking place in the video. The system should also be able to generate captions that are relevant to the military context, such as identifying specific objects, actions, and events that are relevant to military operations. We were also tasked to ensure that the system is able to generate captions in real-time, so that it can be used in live military operations.

We named our final model DeepSub, a fine-tuned vision transformer model with InternVL2 as the base open source model. With our evaluation, our new model has achieved about 4x improvement in military object specificity when generating captions.

An AI-powered web app that combines a curated FAQ system with a real-time chatbot, designed to handle complex topics respectfully and in line with the client's values.

Our team partnered with a client promoting interfaith, interracial, and intercultural education to support trainers facing burnout from answering repetitive, sensitive questions. We developed an AI-powered web app that combines a curated FAQ system with a real-time chatbot, designed to handle complex topics respectfully and in line with the client's values.The platform supports session-based customization, letting trainers tailor content and control access. A human-in-the-loop feedback system allows flagged responses to be reviewed and reused for few-shot training, enabling continuous improvement of the chatbot over time.