Hi, I'm April! I'm a research engineer at NVIDIA, working on multimodal understanding and agentic search, currently with a domain focus in autonomous vehicles but in general interested in Physical AI.
Outside work, I enjoy photography, traveling, driving with music, and spending time with my three cats.
We have taken VLM-based video search from prototype to production: models caption driving clips, and the captions become embeddings for semantic retrieval over a petabyte-scale video datalake. We build the retrieval foundation to support data flywheel that powers end to end training loops.
We post-train video VLMs for driving-scene understanding. My focus is the evaluation side: metrics and benchmarks covering caption quality, temporal reasoning, retrieval behavior, and how well caption embeddings separate driving scenarios, so we can compare fine-tuned checkpoints with confidence.
We're building an agentic search-refinement loop where a rewriter reformulates the query, a search agent retrieves candidates, a judge verifies them, and a refiner sharpens the query until precision converges. It replaces a curation process that used to take hours of manual search and review per scenario, with an evaluation harness that keeps the agents measurable, reproducible, and self-improving.
Small VLMs caption the scene, road entities, and key driving actions in parallel, then consolidate everything into one structured caption, curbing the compounding hallucinations and cost of large-model captioning and improving top-10 retrieval precision by up to 11 points.
Paper (CVF Open Access) · DataMFM Workshop · * Equal contribution
An open-source system for large-scale video data curation that fuses complementary retrieval signals over tens of millions of clips, closing the loop from model failure discovery to training and evaluation.
Coming soon
Vision-language model post-training and evaluation, agentic multimodal search, and large-scale data curation for autonomous vehicles
Multi-agent RAG system for AV engineering Q&A, test generation, and code generation, deployed on Kubernetes and used by thousands of engineers
APAC Corporate Finance Group
CICC Wealth Management
Consumer Credit Business Group
Real Estate Product and Strategy Analysis Division
