PuppyGraph Raises $10 Million to Connect Enterprise Data for AI Agents
- Cargo
- Machine Learning
- Shipped
- Shipper
- Duncan Riley
- Readers
- 841
PuppyGraph says its graph engine lets agents query relationships across existing warehouse and lakehouse tables, with results in under a second across 3,000 tables.

PuppyGraph’s pitch is that enterprise AI agents can reason across connected data without first copying it into a dedicated graph database. The Santa Clara startup has raised $10 million to expand its enterprise sales operation and engineering work, following the general release of PuppyGraph 1.0. It says agents using the software can query across 3,000 enterprise tables and return graph results in less than a second.
That performance claim addresses a familiar scaling problem for agent projects: a pilot can work across a small, curated set of data, then struggle when the task spans many systems and relationships. PuppyGraph says proof-of-concept deployments often cover 30 to 50 tables. At larger scale, descriptions of the data schema can consume an agent’s context window, while queries that traverse several relationships may stall or time out.
Querying relationships without relocating data
PuppyGraph presents its engine as a way to treat data in existing warehouses and lakehouses as one graph. A graph represents entities and the relationships among them, which can make questions involving paths across systems easier to express than a series of disconnected table lookups. The company’s approach lets teams define those connections over data they already store, rather than migrating that data into a separate graph database.
The distinction matters for deployment effort. As Coinbase staff software engineer Xinyu Liu put it, connecting directly to tables already held in Snowflake and Databricks means “standing up a new graph is a schema definition, not a data migration.” That framing suggests the setup work shifts toward describing how tables and their records relate, while the underlying data stays in its existing systems.
For PuppyGraph 1.0, teams begin with a schema drafted by an AI-assisted tool from their tables. People review and approve that schema before users ask questions in plain English. The schema acts as a map of the relationships the graph can traverse; it is also a boundary on what the agent can ask the engine to do.
Schema boundaries and access controls
PuppyGraph includes its own agent for turning questions into graph queries, and it supports connecting other AI tools through the Model Context Protocol. The company says agents cannot invent relationships to fill gaps in the schema: if a query refers to a relationship that has not been defined, the engine rejects it and returns an error written to help the agent respond.
That behavior can reduce one source of unreliable results: an agent presenting a plausible answer based on a relationship that does not exist in the data model. It also means the usefulness of a query depends on the schema’s coverage and accuracy. Teams need to define the relevant connections and keep them aligned with their data; an omitted relationship can block a valid question, while a mistaken one can send queries along the wrong path.
Access controls are another practical concern when a graph query crosses departmental or system boundaries. PuppyGraph says its Snowflake and Databricks connections can use short-lived credentials tied to the user’s identity, avoiding the need to provide long-lived keys. It also says row-level permissions configured in the underlying data remain in effect, with access checked at each hop in a traversal. That per-hop check is relevant because a single graph query can touch multiple records as it follows relationships.
The company illustrates the intended use with a security question: which critical systems could be reached from a compromised account? Answering it may require tracing identities and permissions across thousands of tables in several systems. The example points to a class of questions that dashboards may not cover in advance, though the result still depends on the data connected to the graph and the relationships teams have defined.
Customers, growth and product access
PuppyGraph named Advanced Micro Devices and Palo Alto Networks among its customers, and also identified Coinbase, Datadog and Trend Micro. Chief Executive Weimo Liu, a former Google software engineer, said users will expect questions to cross departments and systems beyond the workflows already encoded in dashboards. Amy Yin, a partner at defy.vc, made a related distinction: an agent may be able to write a query, but executing it across billions of relationships still takes substantial engineering.
The startup says its annual recurring revenue increased sevenfold over the past year, but it did not disclose the starting or ending revenue figure. Particle Future and defy.vc led the $10 million Series A. It was defy.vc’s second time leading a PuppyGraph round, following a $5 million seed round in November 2024. The company says it will use the new funding for sales and engineering. Total funding now stands at $15 million.
Developers can download a free version with a Docker command and no registration, and PuppyGraph says a first query can run about 10 minutes after deployment. That offers teams a quick way to assess whether their tables can support a useful graph schema. The next measure to watch is how the system performs on real workloads: the company’s under-one-second claim and 3,000-table scale describe its stated capability, while the source provides no benchmark methodology or workload details for comparing that result across deployments.