Systems built and measured at production scale.

The analytics team needed to explore time-series data at a scale where the usual answer — pre-aggregate everything and accept the loss of resolution — destroyed the signal they were looking for.

We built a high-density visualization tool combining Rust-based downsampling with Databricks SQL and Apache Arrow zero-copy memory buffers. Queries adapt to the analyst’s zoom level: aggregated views by default, with SQL pushdown retrieving original high-resolution data on demand. Larger-than-memory datasets stay interactive instead of being flattened into something smaller and less useful.

“Petabyte Pitstops with Mercedes, Databricks SQL and Plotly Resampler” · Data + AI Summit 2024

Consulting on agents taught us where they fail: context. Retrieval fetches fragments per session and forgets; the working knowledge an agent needs is a structure, not a search result.

So we built lakecode — an AI coding assistant whose context is compiled into a persistent memory substrate instead of re-fetched every session. The same context-engineering argument as the knowledge-graph study, applied to code: what the agent knows accumulates, survives across sessions, and is governed like any other data asset. Exposed over MCP, so teams keep their own agent front-ends.

the product, public · lakecode.ai

An ongoing research program measuring compute economics and agent architecture across Databricks and Snowflake — the questions platform teams actually decide on, answered with standard suites run under controlled conditions.

Every study ships with the notebooks, worksheets, or prompt lists needed to check the work, including the findings that cut against the obvious answer. When Gen2 warehouses lost to Gen1 at large sizes, that went in the piece.

seven studies, Jan–Jul 2026 · read the research

The supply planning team tracked product ship dates through a manual spreadsheet process — extracts passed between people, with no auditable record of what changed or why.

We replaced it with a Databricks SQL–backed application: an editable AG Grid with SQLAlchemy ORM write-back, so planners keep the workflow they trust while the numbers live somewhere auditable. The published write-up puts the workflow at 60+ steps before and fewer than 10 after.

written up by Plotly · feb 2023