Research Scientist (Jan 2025 - Present)
Designed and constructed an LLM-based system to analyze and detect comment sentiments on Facebook, generated training labels at scale through automated LLM scoring pipelines covering hundreds of millions of comments. Built and deployed end-to-end pipelines with multi-task, multimodal machine learning models to capture fine-grained sentiment nuances, addressing LLM capacity constraints and the enormous scale of Facebook traffic. Conducted controlled A/B testing and integrated models into Facebook's comment ranking system, increasing authenticity by 4.5% and reducing bad vibe prevalence by 22.5%, resulting in significant gains in engagements metrics. Designed and built autonomous ML modeling agent that proposes, trains, evaluates, and iterates on comment ranking models given the goal and design, including subagent delegation, error recovery, and data backed decision making.
ML Software Engineer PhD Intern (May 2024 - Aug 2024)
Assisted the Modern Recommendation System AI Team in building multimodal content understanding models, with a focus on audio understanding and representations. Implemented state-of-the-art models for audio understanding with tokenization, quantization methods and vision transformer-based audio encoder to produce audio embeddings. Built a full pipeline for training and evaluating the model in audio acoustic events prediction task, and in multimodal fusion with visual and text inputs. Constructed a well-balanced internal dataset for audio pretraining, leading to an improvement of ~20% in internal audio understanding evaluation tasks.