Tengyu Ma is MongoDB’s chief AI scientist, a Stanford computer science professor, and the co-founder and former chief executive of Voyage AI, which MongoDB acquired in February 2025. He builds the retrieval systems that allow AI applications to find relevant proprietary information without memorizing entire corporate knowledge bases or processing them afresh for every question.
At Princeton, Ma worked with Sanjeev Arora on the mathematics of non-convex optimization, applying theoretical guarantees to problems including sparse coding, topic models, word embeddings, and dynamical systems. His dissertation received an honorable mention for the ACM Doctoral Dissertation Award; his subsequent distinctions include a 2018 Conference on Learning Theory best-paper award and a 2021 Sloan Research Fellowship.
At Stanford, his research expanded into deep learning, representation learning, reinforcement learning, foundation models, robustness, and high-dimensional statistics. He teaches machine learning and advises doctoral researchers; Hong Liu, a member of his research community, became a Voyage AI co-founder.
Voyage developed embedding and reranking models that help enterprise applications identify useful documents before generating answers. MongoDB’s acquisition brought those capabilities closer to the databases holding organizations’ operational information. Ma subsequently became MongoDB’s chief AI scientist.
- Retrieval as durable infrastructure. Ma considers retrieval-augmented generation a lasting architecture for enterprise AI: proprietary information remains externally governed, updateable, and permission-aware, while models receive only material relevant to each request. Fine-tuning complicates deletion and access control; sending entire document collections into long context windows increases cost. His AI Engineer talk frames retrieval as essential infrastructure, not a temporary workaround.
- Context without bloated prompts. Small document chunks improve retrieval precision but can lose surrounding information. Contextualized document embeddings preserve both passage-level meaning and broader document context; Voyage’s subsequent model integrated automatic chunking and support for longer documents.
- Efficiency built into representations. Matryoshka learning and quantization-aware training reduce vector dimensions or numerical precision while managing accuracy losses. Specialized code embeddings can improve retrieval quality and make compression more effective for domain-specific collections.
- Search beyond extracted text. Multimodal document retrieval can embed screenshots of PDFs, slides, tables, and video frames directly, preserving visual structure and reducing fragile parsing pipelines. Ma expects stronger retrieval models to absorb more preprocessing, contextualization, and orchestration currently handled by application developers.
