data-agents
(5)
Seek AI Taught Databases to Answer Back
It gave curious colleagues a way around the analytics ticket queue, while keeping SQL and human review close at hand. Then IBM came calling.
Luca Martial Is Teaching AI Agents to Ask What the Numbers Mean
Before Luca Martial built a context layer for data agents, he studied bodies, cities, clinical language, and model failure. The through-line is less about AI than about making complicated systems answerable to people.
Kaelio
Kaelio is a San Francisco startup (Y Combinator, 2025) building the open-source context layer that makes AI data agents reliable. Its core project, ktx, auto-builds a governed layer of business knowledge and executable metric definitions from a company's data stack - Markdown wiki pages plus YAML that declares tables, joins, grain, measures and filters. Instead of an agent writing raw SQL that looks right but returns wrong numbers, it asks ktx for a metric and ktx plans the joins and compiles warehouse SQL. Kaelio ships ktx (Apache 2.0), a hosted ktx Cloud with governance and observability, and a managed Data Agent that answers questions in Slack, web and email with source-backed results.
Rohan Kodialam
Rohan Kodialam is the co-founder and CEO of Sphinx, a New York applied-AI research firm building agents that actually understand data. A former Citadel quant who learned to squeeze alpha out of messy alternative datasets, and an MIT-trained researcher who once taught transformers to read hierarchical data, he started Sphinx on a simple observation: software engineers got Copilot, everyone else got ChatGPT, and data scientists got nothing. In 2025 Sphinx launched out of stealth with a $9.5M seed led by Lightspeed and backers including Steve Cohen and Naveen Rao, betting that the next frontier of AI is not language or code, but the spreadsheets and warehouses where real decisions get made.
Jason Cui
Jason Cui is a Partner at Andreessen Horowitz (a16z) investing at the intersection of foundational AI, developer tooling, and scientific platforms. A Harvard computer science graduate, he co-founded Jemi (acquired by Brat TV in 2023), led product at Databricks, and previously worked at Uber and Hulu. Named to Forbes 30 Under 30 in 2021, he joined a16z to back the infrastructure layer of the AI era, with bets like the $150M seed round for Inferact, the next-generation commercial inference engine built by the vLLM maintainers.