We build the full stack from data systems to autonomous agents. Our engines make data ultrafast to query; our AI databases let people express intent in natural language; our agents ground large language models in enterprise data across text, tables, images, and graphs; and our embodied agents carry the same ideas into robots that act in the physical world.
Enterprise knowledge is scattered across text, tables, images, and knowledge graphs. We build retrieval-augmented agents that reason over all of these modalities at once: multi-granular retrievers that fuse table and text evidence, late-interaction retrievers that scale to open-domain multimodal corpora, and agentic traversal strategies that learn from their own failures. We also design principled benchmarks so that progress on complex multi-hop question answering can be measured rigorously.
We are redefining what a database engine is in the era of LLMs. CADENZA compiles natural-language intent into task-specific operator DAGs, turning semantic queries into optimizable plans instead of monolithic LLM calls. TurboLynx is the world’s fastest embedded graph DBMS—a schemaless engine for general-purpose analytics over graphs and tables—and Themis executes relational queries on GPUs. Underneath, we keep pushing the core of query optimization: learned cardinality estimation with theoretical guarantees, cost-efficient batching of massive small-query workloads, and cache replacement for direct-access cloud OLAP.
Vision-language-action (VLA) models let robots follow open-ended language instructions, but they struggle with what makes a task yours—”bring my cup.” We study how to personalize and ground VLA policies with lightweight visual prompting and in-context adaptation, without retraining the underlying model, and how to connect embodied agents to the knowledge and retrieval infrastructure we build for enterprise agents.
Serving and reasoning with LLMs is, at its core, a systems problem: memory hierarchies, scheduling, and cost-aware planning. We are bringing our database-engine expertise—query optimization, caching, and workload-aware execution—to LLM inference, and extending it to multi-step reasoning so that agentic workloads run faster and cheaper without sacrificing answer quality. This is where our engines, AI databases, and agents converge, and it is the next direction of the lab.
See our full publication list.
