I won second place in an international competition to detect harmful content in multimodal and multilingual memes. The competition is done in a zero-shot fashion, with no training or validation data released. I built a reusable pipeline that can robustly evaluate memes in under 2 seconds, and a QLoRA fine-tuned LLM, trained on a synthetically generated dataset, with BLIP captioning and multilingual OCR. Competition report is published as a short paper in ACM Web Conference 2024 proceedings.

I led a collaboration project with Singapore's Ministry of Education to implement and benchmark LLMs, and evaluate open-ended language tasks to create solutions to assist educators and expand student access to feedback. This led to a successful implementation on AWS, and also a research paper that has been accepted to NAACL 2024. Also made open-source contributions to popular LLM benchmark library Alpaca-Eval.

I led a collaboration project with Singapore's Government Technology Agency to integrate generative AI capabilities into GovTech's ML/AI platform called MAESTRO. Through the project, we collaborated on inference optimizations for large language models, achieving 75% cost reduction in inference costs with no significant decrease in quality benchmarks. Public sector users develop on the platform to enable diverse and impactful use cases; for example, CPF Board is using the platform for call transcript summarization for over 600,000 call transcripts per year and Ministry of Manpower developed a LLM-powered sensemaking tool to process over 1 million documents and save over 2,000 work hours annually.

I led the engineering efforts to integrate LLMs for regional languages in SageMaker Jumpstart. This includes SEA-LION, a 3B-7B model trained on Southeast Asian languages. These include making changes to popular optimized LLM inference libraries like Text Generation Inference.

I led the science and research work to enable immersive storytelling with LLMs and diffusion models. I worked on decreasing overall latency from 5 mins to under 30 seconds, measuring and enhancing art style and character consistency, and reducing occurences of undesirable content. The immersive storytelling was an overwhelming success for National Libray Board, resulting in nationwide news coverage and repeated extensions of the project and display.

Working closely with stakeholders in National Parks Board, my team and I developed a cloud-based crowd estimation system. By connecting over 180 cameras nationwide to our AI-enabled people counting backend, we were able to reduce and optimise the heavy operational requirements for NParks to deploy their offices to ensure park safety, and also provided members of the public with useful real-time data to self-manage their own itineraries. This event-driven system was delivered in an iterative approach, scaling quickly from our initial pilot of 5 to 100+ cameras within 3 months. The public-facing website receives over 450k-800k monthly visits since 2020, with the project receiving media coverage by Singapore primetime news.

Treepedia measures the canopy cover in cities. We've developed a scalable and universally applicable method by analyzing the amount of green perceived while walking down the street. The visualization maps street-level perception only, so your favorite parks aren't included. Our work has been featured on popular news outlets such as Wall Street Journal, The Guardian, Forbes and Wired. Our Treepedia 2.0 paper is accepted at the IEEE BigData Congress 2018, and we explore our deep learning application in the area of climate change in a NeurIPS 2019 Climate Change workshop paper.