Saurav Joshi

I’m an AI Scientist at Spine AI (YC S23), where I lead the AI/ML research behind a stateful, general-purpose multi-agent system for long-running missions. A lot of my work is building the core product, including the agentic framework, from the ground up. Much of it is eval-driven — I design evaluation harnesses and benchmarking pipelines across agentic benchmarks like Terminal-Bench, DSQA, and GAIA, where we reached state of the art, outperforming agentic systems from frontier labs like OpenAI and Anthropic. I run systematic ablations and failure-mode analyses to sharpen agent reasoning, planning, and tool-use on long-horizon tasks. I also architect the multi-agent orchestration itself: hierarchical planning with multimodal context passing and long-context management, on a Redis-backed DAG execution engine for parallel tool execution and distributed scheduling.

Before Spine AI, I graduated from the University of Southern California with a Master’s in Data Science, where I spent two years at the USC Information Sciences Institute doing research at the intersection of machine learning and knowledge graphs, advised by Prof. Filip Ilievski and Prof. Jay Pujara. My work spanned natural language processing and social media — including retrieval-augmented generation and long-context question answering — as well as a DARPA collaboration between USC ISI and ICT on knowledge management and granular knowledge retrieval. Broadly, I’m drawn to building explainable, efficient agents for real-world tasks.

Alongside my studies, I contributed to Google Summer of Code twice: in 2023 on question-answering over DBpedia with pretrained auto-regressive models (continued pretraining and fine-tuning of StarCoder-1B), and in 2022 on template discovery for neural question-answering over DBpedia, translating natural-language questions into SPARQL.