An AI agent investigating a compromised account may need to follow a trail across thousands of enterprise tables to determine what an attacker can reach. If a database query takes half an hour, the investigation stalls. The agent cannot decide what to examine next until the company’s systems finish answering its last question.
Weimo Liu, cofounder and CEO of PuppyGraph, sees that bottleneck as an opening in the enterprise AI market. Companies have spent years collecting data, but tracing relationships across it can require cumbersome queries or copying information into a separate database. PuppyGraph lets agents explore those connections as a graph while the underlying records remain in their existing systems.
The company recently announced a $10 million Series A, led by Particle Future and defy.vc, bringing total funding to $15 million. PuppyGraph reports sevenfold growth in annual recurring revenue over the past year, although it has not disclosed the revenue base. Customers include AMD and Coinbase, alongside cybersecurity companies such as Palo Alto Networks and Trend Micro. Datadog and Netskope also use the platform.
Liu says some buyers turn to PuppyGraph after struggling to deploy AI systems that translate natural-language questions into SQL, the language commonly used to query databases. These companies often begin by building their own text-to-SQL engines, investing heavily in extensive context layers and incorporating historical SQL queries to improve accuracy. Despite these efforts, many still struggle to move their systems into production.
“Eventually, these companies begin exploring alternative approaches, researching ontologies and knowledge graphs, which is often how they discover us. When they try PuppyGraph with their enterprise data, including sales tables, the experience can be quite different. Building an ontology graph on top of that data allows the system to work without significant additional effort,” he tells me. “Once the ontology graph is in place, users can begin asking questions directly against their existing data.”
Liu says the distinction will become increasingly important as agents take on more complex investigations. “Access to enterprise information alone is insufficient if the underlying query engine cannot retrieve answers quickly enough for agents to determine their next steps and continue their work.”
Inside PuppyGraph’s database-free graph analytics
Before founding PuppyGraph, Liu worked as a software engineer and research scientist at TigerGraph, then spent more than four years at Google working on GoogleSQL and the F1 query engine. At TigerGraph, he saw how preparing data for graph analysis could become a substantial project of its own. He described one large bank spending 18 months loading its data into the system.
PuppyGraph extends that approach across existing enterprise data sources, allowing users to explore relationships as a single graph without moving or consolidating the underlying records.
The platform connects to systems including Snowflake and Databricks and supports open table formats such as Apache Iceberg. Liu explained that this separation is fundamental to its performance. “The important question is what trade-offs you are willing to make, because achieving all those benefits inevitably means giving something up. In our architecture, we have deliberately chosen to forgo transactional updates, including ACID transactions. That is a conscious architectural decision,” he says.
To explain the reasoning, Liu drew a parallel with ClickHouse, a database optimized for online analytical processing (OLAP), and PostgreSQL, which is widely used for online transaction processing (OLTP).
“People do not choose ClickHouse simply because it is a better database than PostgreSQL. They choose it because ClickHouse is specifically designed for OLAP workloads, whereas PostgreSQL is designed for OLTP workloads. Each system is optimized for a different purpose,” he says. “Our ambition is to become the ClickHouse of graph technology. We want to deliver the best possible performance for complex graph analytics, and the trade-off is that we deliberately give up OLTP capabilities.”
Why AI agents need more than SQL
PuppyGraph 1.0 builds on that foundation by proposing graph schemas from existing enterprise tables and allowing users to ask natural-language questions through its built-in AI agent. External tools can also connect through the Model Context Protocol (MCP).
Asked about competing ontology initiatives at larger platforms, including Databricks, Liu argues that their reliance on SQL continues to shape how they approach enterprise AI. “These companies are trying to build what they call data agents, but in many cases, those systems still function more like data copilots. You provide a natural-language question, and the system translates it into a single SQL statement,” he explains. “Along the way, vendors realize that generating accurate SQL requires an ontology layer to provide additional context about the underlying data. Without that layer, achieving reliable accuracy becomes extremely difficult.”
Liu argues that an ontology should be part of the architecture from the outset and that its underlying relationships are better queried directly through a graph query language. “We believe our architecture is more agent-native and much closer to what an autonomous agent is supposed to be, rather than simply another copilot. An agent needs a harness and auditing capabilities outside the large language model to help ensure accuracy. The answer is not simply to make the language model better at generating SQL.”
He points to cybersecurity investigations that can run for 10 to 20 minutes and generate roughly 10 pages of findings, with each query uncovering evidence that informs the agent’s next step. To keep these investigations grounded in the underlying data, PuppyGraph rejects references to relationships that have not been defined in the graph and returns corrective feedback. Likewise, the platform examines whether identifiers are unique and how records match across different fields, expanding its sampling when the available evidence is insufficient. Enterprise data teams must still review and approve the resulting graph against their business logic.
Access control presents another challenge as agents traverse increasingly complex relationships. PuppyGraph says it enforces row-level permissions throughout the traversal process, including intermediate steps that might otherwise expose restricted information.
Can AI agents query enterprise data in real time?
PuppyGraph reports completing a 10-hop query across billions of connections in 2.26 seconds on a four-machine cluster. Liu says the benchmark used a data lake, whose metadata let the engine skip irrelevant information without adding its own indexes or cache. Operational databases require different accommodations because additional reads can burden systems already handling production work.
“We have two approaches to handling this category of data sources. The first is lazy fetching, where we retrieve the necessary data only when a query requires it. The second approach involves caching selected data within our own system, allowing us to execute queries more quickly,” Liu explained. “Caching also helps reduce the read workload on the underlying database. Systems such as PostgreSQL and MongoDB often already handle substantial workloads, and we do not want to introduce additional read traffic that places even greater pressure on them.”
Avoiding a separate graph database therefore does not mean the engine never holds source data. However, Liu did not disclose production cost figures or latency distributions under competing workloads.
Customer deployments offer a clearer picture of the practical benefits. At Palo Alto Networks, the value extends beyond faster queries to reducing the engineering work required to maintain applications. According to PuppyGraph, the cybersecurity company’s user-facing application can traverse multiple relationships across its existing data platform. That architecture eliminates the need for separate data-loading pipelines that its engineering team would otherwise have to maintain, while enabling users to perform graph analytics in real time.
AMD uses PuppyGraph to connect engineering information that would otherwise remain scattered across systems. By querying connections directly from Apache Iceberg tables, AMD avoids maintaining separate graph-data stores and the pipelines required to populate them. PuppyGraph reports that the approach reduced query times from minutes to fractions of a second across millions of relationships.
Native graph engines vs. SQL-based approaches
Neo4j’s Virtual Graph challenges the idea that avoiding data migration alone distinguishes PuppyGraph. Meanwhile, Google’s BigQuery Graph and Microsoft’s Fabric graph capabilities give customers additional reasons to consider their existing suppliers before adopting a specialized platform.
Liu characterizes Neo4j’s approach as translating Cypher, a graph query language, into SQL for execution by the underlying data warehouse.
PuppyGraph instead uses its own distributed graph engine, making its proposition more specific than simply providing an ontology or eliminating the need for a separate database. Graph capabilities built into individual platforms may be sufficient when the relevant information resides within those systems. PuppyGraph, by contrast, seeks to operate across them while leaving the underlying records in place. Its opportunity therefore depends partly on how much analytical work enterprises need to perform beyond the boundaries of any single supplier.
“That is why having a native graph engine is so important. The fundamental advantage is that we can optimize graph workloads specifically for the demands of graph processing, rather than relying on a SQL engine as a workaround. This allows us to optimize graph query execution directly and deliver substantially better performance for the types of complex queries our customers need to run,” Liu says. “If you simply translate Cypher into SQL and execute it on Snowflake or another database, you are essentially providing an alternative syntax for SQL.”
Liu asserts that this distinction becomes particularly important for AI agents, which generate queries based on the task at hand without necessarily accounting for their complexity or the computational demands of executing them. “As a result, the underlying graph engine must be capable of executing highly complex queries efficiently, even when those queries are generated dynamically and their complexity cannot be anticipated in advance.”
Still, performance claims require workload-specific comparisons to establish how much of an advantage PuppyGraph can deliver. Neo4j’s embrace of virtual graphs helps validate the premise that graph analytics does not necessarily require dedicated graph storage, allowing PuppyGraph to spend less time defending its architectural approach and more time demonstrating its performance.
Winning enterprise trust in a crowded market
Customer demands are expanding PuppyGraph’s reach. Liu says it recently added SAP support for supply chain agents and operational metrics support for observability agents.
Liu says PuppyGraph is in late-stage talks with Goldman Sachs and Morgan Stanley, though he does not confirm either as a customer. He also recalls a CrowdStrike principal architect asking detailed questions about the platform’s capabilities before remarking that the team must already have worked extensively with companies building cybersecurity agents.
“These are some of the largest enterprises in the world, and they are generally reluctant to take unnecessary risks. They would not ordinarily choose a product from a Series A or Series C startup if they could solve the same problem with technology from a more established vendor,” he explained. “So when these enterprises ultimately decide to work with us, it is often because they have already evaluated other approaches and discovered that those solutions could not deliver the results they needed.”
THE CONTROL LAYER VERDICT
PuppyGraph’s strongest argument is that enterprises should be able to expand what their AI agents can investigate without rebuilding the infrastructure holding their data. Its ability to query relationships across existing platforms gives it a credible opening, particularly when an investigation extends beyond one supplier’s systems. The commercial test is whether that advantage remains compelling enough to justify a separate vendor as larger competitors improve. If Liu can sustain reliable performance as workloads grow, PuppyGraph could earn a durable role in enterprise AI. That will require reducing the work needed to reach a trustworthy answer, including the human effort required to keep the underlying relationships accurate.
