Chalk

?? Chalk term of the day: Resolvers ?? Resolvers determine how features are computed and retrieved at inference time from data sources such as Postgres, Iceberg, and 3rd party clients. You can implement Resolvers using SQL, pure Python, or inline expressions (lambda-like). ???SQL Resolvers: Best for structured data like attributes from a Postgres table. Define SQL resolvers with familiar syntax but with a .chalk.sql file extension or the make_sql_file_resolver function. We only fetch the exact data needed for each inference by pushing down filters end-to-end. ???Chalk Expressions: Low-latency, inline computations guaranteed to be accelerated (C++). ???Python Resolvers: Ideal for complex logic, API calls, or leveraging external libraries. We parse each Python resolver's AST (even numpy functions) and accelerate scalar transformations by composing them into Chalk Expressions. Check out our demo video to see Resolvers in action!

Pi is irrational. So is expecting real-time ML to scale without the right infrastructure. Happy Pi day from Chalk!

At Chalk, we've made it possible to iterate on your features from a Jupyter notebook ?? Easily import your team’s features and classes in just one line of code and start prototyping models that you can share with teammates over a branch. In the code example below, load_features pulls feature classes from the feature store directly into your notebook namespace. Add new features and start querying away!

Melanie Chen started a Chalk book club! In every session, someone presents a paper or book. Topics include: - DeepSeek-R1 paper (pictured: talk led by Sai Atmakuri) - 3D Gaussian Splatting for photorealistic image rendering - Andy Pavlo's database roundup on the "Databricks vs Snowflake gangwar" - Monolith, the real-time ML infrastructure behind TikTok - Google Spanner paper on building a globally distributed database I think surprise is conserved – discoveries in one field often transform and reappear in another. I've found that being curious about different domains and learning unexpected things often helps us solve problems in our work. If you're looking for a community where you can nerd out over interesting papers, Chalk is hiring! P.S. Nathan Fenner did a truly great presentation on ellipsoid "splats" vs. traditional voxels - slides in the comments.

Discover how LLMs are being used to prevent sophisticated social engineering attacks in real-time ?? Join us with Doppel on March 27th for a live demo, moderated by Elliot Marx, co-founder at Chalk where we’ll share how Doppel: 1/ Synthesizes structured metadata with LLM outputs to improve detection accuracy 2/ Builds scalable feature retrieval systems for real-time use cases 3/ Optimizes for faster computational performance and development velocity Don't miss this deep dive into real-time hybrid LLM pipelines. Whether you're fighting fraud, assessing risk, or driving AI in mission-critical operations, you'll leave with a blueprint for integrating structured and unstructured data into production ML systems. Sign up here https://lu.ma/1176k2v9

No one has to read docs for Instagram because it's intuitive by design. That ease-of-use is a key difference between consumer and enterprise products. You open the Instagram app and it works. On the flip side, IBM mainframes are unusable without expensive, high-touch expert support. I aspire to make Chalk’s developer experience feel more like the intuitive feeling of a consumer app. That's why we built everything in Chalk to feel like the Python tools engineers already know instead of another proprietary system they have to learn from scratch. We've all heard the joke "no one reads the docs" – and in product development, that should be the goal, not a user problem.

Chalk's MLE function is our take on Palantir's Forward Deployed Engineer model. MLEs at Chalk have one primary mission: to help customers succeed. They handle critical support work, such as training sessions and guiding customers to go-live. While not the primary goal, this role drives most of the product roadmap. By working closely with customers, they identify gaps in the product and discover what people need. If a customer needs a tool to solve a problem, we will build that tool. For example, Samuel Mignot realized that data scientists weren’t using the same codebase as the engineering team. In response, he created a method to load code over the network directly into a notebook, removing the need for the engineering codebase. If you're an engineer who loves working with customers and building products, let's chat!

elvis kahoro spun up a demo on how Chalk's windowed aggregations provide a clean way to combine real-time updates with historical data ?? Chalk makes it easy for engineers to roll existing data from Postgres / Iceberg / data warehouse into time-based buckets, while also dynamically integrating live data. Materialize and tile aggregations in just a single line of code — no extra pipeline work required! P.S. This dramatically accelerates feature computations without sacrificing accuracy ??

When your startup is growing 3x, scaling support becomes a real challenge. The volume of inbound messages makes AI customer support bots an inevitable temptation. However, it’s hard to use them when your customers are ML engineers who can spot deflection tactics immediately. I think the best approach for AI support tools in our space is straightforward self-awareness, not warm human-sounding messages that pretend to be Rebecca from Customer Success. The ideal bot says "Hello, I'm an annoying bot. A person will review this thread shortly, but here are my best ideas about how to help you while we catch up on the context." It's important to acknowledge the shortcomings of the technology in order to make the best out of it, rather than pulling wool over customers' eyes.

When lunch break turns into friendly competition. Who will emerge as the ace at Chalk? ??