AI/ML
Startup PuppyGraph has tech to build graph-based relationships from database tables to help AI agents answer high-level queries.
A business exec or manager might want to know if a particular customer account is compromised, or which critical systems an attacker could reach. If they ask an agent to answer this then the agent has to follow relationships across identities, permissions, and infrastructure. This might require following connections across 3,000 or more database tables and multiple systems. As the number of tables grows, schema descriptions consume more of an agent's context window and the possible join paths multiply. Deep relationship queries can become too slow for interactive agent investigation, or time out entirely.
Weimo Liu, cofounder and CEO of PuppyGraph, said: ”The insights of enterprise data will soon be generated by agents. Users will expect to ask questions that cross departments and systems, beyond the workflows someone built into a dashboard. That requires agents to navigate thousands of tables, follow the right relationships, and retrieve evidence in subseconds.”
The company has released its v1.0 software and raised a $10 million A-round to fund engineering and enterprise go-to-market expansion led by Particle Future and defy.vc, bringing total funding to $15 million.
PuppyGraph’s technology enables an AI-assisted graph schema builder that reduces manual modeling by automatically generating a graph schema from existing tables for teams to review and approve. Envisage a graph query layer on top of the warehouses and lakehouses you already run. Users can investigate connected enterprise data in plain English through PuppyGraph’s built-in AI agent or compatible AI tools they already use, connected through MCP.
The query engine rejects undefined relationship references and returns feedback that agents can use to correct their queries, and so avoid errors from guessing how data connects.
PuppyGraph brings graph compute directly to Apache Iceberg, Databricks, Snowflake, ClickHouse, DuckDB, PostgreSQL, Microsoft Fabric, Apache Hudi, and other supported sources. Its distributed engine supports petabyte-scale and subsecond graph queries. PuppyGraph completed a 10-hop query across billions of edges in 2.26 seconds in a four-machine cluster.
Agents gain graph compute alongside SQL compute over the same data. Applications choose the engine that fits each question without maintaining a separate graph copy. That would require an ETL process and separate graph database which Puppygraph's tech avoids.
PuppyGraph 1.0 enforces configured row-level permissions throughout graph traversals, including intermediate steps. Single sign-on, role-based controls, and user impersonation tie data access directly to user identity. Credential vending enables PuppyGraph to securely obtain and use short-lived, identity-aware credentials when connecting to data platforms such as Snowflake and Databricks, without requiring users to expose or manage long-lived credentials. Administrative audit logs record management activity, providing greater visibility and accountability across the entire data access path.
Developers can start free with a Docker command, with no form or payment required, and go from deployment to query in 10 minutes. Find out more here.
Bootnote
PuppyGraph customers include AMD, Palo Alto Networks, Datadog, Netskope, Coinbase, Trend Micro, Baker Hughes, Sola Security, HighStreet Insurance Partners, and Prevalent AI. Its annual recurring revenue grew 7x over the past year.
Ran Xia, Engineering Director, Palo Alto Networks, said: “For a user-facing application, query latency and data freshness are non-negotiable. PuppyGraph allows us to execute multi-hop graph traversals directly on our existing data platform without moving or duplicating a single row. By removing the ETL bottleneck, we eliminated pipeline maintenance entirely while delivering lightning-fast graph analytics to our end users in real time.”