Request a demo
Interview

How Pometry works: an interview with CTO Dr Ben Steer

· Dr Ben Steer, CTO

Music: “Inspired” by Kevin MacLeod (incompetech.com), licensed under CC BY 4.0

Dr Ben Steer is Pometry’s CTO and co-founder, and sits on the board of directors of the Graph Data Council, which builds benchmarks and standards for the graph industry. We asked him the questions architects and engineers ask us most: how the engine is built, how it handles real-world data, how it keeps AI honest, and what it takes to run.

Lightly edited for reading. Select a time to jump to that answer in the video.

What makes Pometry different?

”Pometry’s primary differentiator is the way that it handles time as a first-order citizen: understanding how entities, their relationships, their communities, and your data ecosystem as a whole change over time.”

Why build a new engine rather than use an existing graph database?

”We built Pometry’s graph engine from the ground up because the systems that were available did not scale to the size of data that we were seeing: tens of terabytes. They didn’t deal with time at all, so you couldn’t roll between previous versions, and you couldn’t have time series analytics as part of your queries.”

What does it mean to have time built in?

”Time built in is a philosophy around making sure that every entity and every relationship has a full chronological log of every change which has occurred to it: every deletion, every insertion, every change of property. This is stored in a manner where we can freely move between different versions and do comparisons between versions, but also, as part of our analysis, we can have network analysis with time series baked in.”

“Every entity and every relationship has a full chronological log of every change which has occurred to it.”

Dr Ben Steer, CTO, Pometry

Why Rust and Apache Arrow?

”We chose Rust for memory safety, for optimisations, and for its integration with a variety of other low-level tools. We chose Apache Arrow because it allows us to store data in a manner where we can zero-copy it into memory, and because it’s the substrate upon which many other tools communicate. So we can load directly from Parquet, from Iceberg, from Databricks and from other databases, and we can push out formats that all of these other tools understand as well, which means we work very nicely in any form of pipeline.”

Why keep the data on disk rather than in memory?

”We keep the data on disk because we don’t want use cases to be limited by the machines that you have available. If you want to load 10 terabytes of data into memory, that requires an absolutely gargantuan box, and for a lot of use cases that just isn’t viable.”

How do you get messy data from dozens of systems into one model?

”We set up Pometry to be able to ingest from multiple sources, be they structured data like Parquet or CSV, or unstructured data, where we can do entity resolution to extract entities and push these into the system.”

How does it handle data that arrives late or out of order?

”Any data that’s ingested into Pometry is sorted into chronological order internally, so if you’re loading from multiple sources, all of which have different timestamps on them, we handle all of that. We also manage event time versus processing time, so you can ask questions into the past and understand what you knew at that point.”

How do you stop AI agents making things up?

”If you have an AI MCP server querying Pometry, we have a variety of tooling that lets you explore the ontology via similarity and expansion in the graph space, and then map that to our underlying queries. So you can allow your agent free rein to ask any question it likes, but constrain that within the bounds of what you have in your model. Any question that comes to Pometry is answered, but the answer also references entities and relationships within the network, meaning that it can be traced.”

“You can allow your agent free rein to ask any question it likes, but constrain that within the bounds of what you have in your model.”

Dr Ben Steer, CTO, Pometry

Does my data leave my environment?

”Pometry is deployed as a single binary. Your data never leaves your ecosystem. You can run it on EC2 instances, completely air-gapped, inside Databricks or Snowflake, or locally on a machine like a laptop.”

How can my technical team build on it?

”If you’d rather build your own tools instead of using our full product suite, our engine has APIs for Python and for user-defined functions in Rust. You can query a Pometry server via GraphQL, Cypher or RDF SPARQL, and you’re also able to box this up inside our MCP server.”

What happens when things go wrong?

”If something were to go wrong with the box Pometry was deployed on, a full redeploy takes less than five minutes. An in-memory system, or a system that has to reload from other systems, would take many hours at the size of data that we deal with.”

How performant is Pometry?

”Pometry was designed from the ground up to deal with tens of terabytes of data, with responses returning in under 100 milliseconds.”

Experience Pometry in action
with a live demo.

Find out how you can get answers faster with unique institutional intelligence.