Artem Keydunov

Artem Keydunov, also credited as Artyom Keydunov, is Cube’s co-founder and CEO. His AI-relevant work centers on semantic layers: executable definitions of business metrics that agents can query rather than reconstructing their meaning from database tables. With co-founder and CTO Pavel Tiunov, he announced the agentic analytics platform D3 in June 2025; his later writing argues that governed execution should sit at the center of an agent’s context layer.[1][2][3][5]

Keydunov and Tiunov jointly announced D3 on June 2, 2025, describing a workflow in which humans and agents could modify each other’s queries and reports. Their architectural argument went beyond supplying metric descriptions to a model: agents should query the semantic layer itself so that execution uses the governed definitions. They favored SQL because it is established, well documented, and familiar to existing models.[3]

Their Semantic SQL design combined shared semantic definitions with ad hoc calculations. Users could review and change generated queries through a visual editor or SQL code, then save calculations for reuse or add them to the shared model. The architecture separated three tasks—building the semantic layer, constructing Semantic SQL, and building visualizations—and incorporated compilation checks, linters, prompts, documentation, and memory retrieval as safeguards. The authors presented this separation as a way to improve reliability.[3]

On October 29, 2025, Keydunov announced general availability of Cube Agentic Analytics, removing the D3 name and bringing semantic-layer and agentic-analytics offerings under one interface and the Cube name. He emphasized fine-grained human control: agents could build and update analysis while users corrected query fields, filters, and presentation details.[4]

In his July 10, 2026 article, Keydunov called a semantic layer “executable context.” He distinguished it from descriptive context—catalogs, glossaries, ontologies, and business documentation—and from memory of earlier decisions. In his proposed architecture, metrics, dimensions, joins, and access policies belong in the semantic runtime; descriptive information is retrieved as needed through MCP, while memory belongs in the agent harness surrounding the model.[5]

His concrete example was MEASURE(completed_percentage): instead of asking an agent to invent the calculation, Cube’s rewrite engine expands a defined measure at the appropriate aggregation level, in the warehouse’s SQL dialect, with access rules compiled into the query. Keydunov argued that permissions should be enforced during compilation, before SQL reaches the warehouse. His recommendation was to build the agent harness around this executable core and assemble the remaining context around it.[5]

The semantic infrastructure behind these later decisions originated in Statsbot, an earlier Slack analytics bot. In his October 2023 Latent Space interview, Keydunov recalled starting Statsbot in 2016, drawing on a project he had built while working at a company making software for schools. He wanted statistics from sources such as New Relic, Google Analytics, and production databases available inside Slack. He published the side project on Reddit and recalled gaining more users after Slack featured it on its application-directory front page.[1]

Keydunov built the first natural-language implementation with regular expressions and credited Pavel Tiunov with subsequently exploring models. He recalled that the system handled isolated requests unevenly and lacked the ability to ask follow-up questions. Statsbot separated interpreting a request into metrics and dimensions from translating that multidimensional query into SQL. Cube supplied the latter framework, allowing customers to define what terms such as “active opportunities” meant for their own datasets.[1][2]

As users found value in that framework beyond Statsbot, Keydunov and his collaborators extracted and open-sourced Cube in 2019, according to his account. He summarized its origin directly: “We built a Cube at Statsbot because we needed it.”[1]

In the same 2023 interview, Keydunov described indexing textual explanations of metrics, dimensions, relationships, and possible dimension values, turning them into embeddings in vector storage, and supplying that context to a model. He argued that the model should request a metric, grouping, and filter from the semantic layer while the layer handles complex warehouse joins and aggregation pitfalls. This was his rationale for reducing opportunities for query errors.[1]

He also advocated treating semantic definitions as code: version-control them, discuss competing metric definitions through pull requests, and retain multiple versions when necessary. His enthusiasm for copilots included practical limits. In October 2023, he said generated transformation and integration code still needed human oversight, and described production analytics applications using additional Python code for calculations outside the model.[1]