February 2026 Summaries

Adrian Brudaru, in his exploration of ontology-driven dimensional modeling, argues that the modern data stack suffers from theoretical limitations, primarily due to over-reliance on dimensional models and semantic layers that fail to capture the complexities of real-world systems. He emphasizes that data modeling should start from a business ontology, which serves as a "blueprint of truth," offering deeper insights and enabling AI to transition from mere reporting to strategic analysis. Ontologies provide a framework that allows for semantic flexibility, meaning-based search, and global interoperability, unlike traditional data models that often lead to misinterpretations when used alone by AI systems. Through an experiment, Brudaru demonstrates that an LLM relying solely on standard data models can misinterpret trends, while one augmented with ontology can accurately assess business contexts and provide strategic recommendations. The article advocates for a shift toward ontology-driven approaches in data modeling to bridge the gap between data representation and the nuanced realities they aim to depict.

Adrian Brudaru's article delves into the intricacies of modeling unstructured data for AI agents, illustrating how traditional structured data concepts can be applied to this domain. It introduces the Canonical Data Model (CDM) as a vendor-agnostic ontology that helps achieve data consistency and decoupling in structured data environments, while dimensional models optimize query speed. The semantic layer has gained importance as it aids large language models (LLMs) by providing a governed context, preventing them from producing inaccurate queries. For unstructured data, knowledge graphs serve as a canonical model, organizing data as nodes and edges to maintain meaning and relationships. The piece highlights Cognee, a Python SDK and knowledge engine, which builds structured memory for AI agents through ingestion, graph construction, graph maintenance, and hybrid retrieval. This system turns scattered data into a single, structured memory system, improving query performance and providing a robust framework for context engineering, which is increasingly adopted by enterprises like Bayer.

In this analysis, a focused evaluation was conducted on the generative models used in a Retrieval-Augmented Generation (RAG) system, revealing significant improvements when newer models like Gemini 3 and GPT-5.2 were employed, compared to a legacy model from 2023. Despite achieving a performance increase from 3 to 10 out of 14 correct answers without altering the retrieval pipeline, persistent failure modes such as "needle-in-a-haystack" retrieval failures, hallucinations in multiple-choice scenarios, and omissions of critical details remained. The study underscored that while upgrading generative models offers substantial gains in handling noisy contexts and simple queries, there is a performance ceiling that suggests further improvements will require enhancements in the retrieval process, specifically through better embedding models to address nuanced documentation differentiation. The findings indicate that while model upgrades are beneficial, they are insufficient alone, pointing to the need for more comprehensive iterations that include retrieval assessment.

The blog post promotes a workshop hosted by DataTalks.Club on AI-assisted data ingestion with dlt, led by Aashish Nair, emphasizing a shift in data engineering from manual script writing to designing intelligent systems that automate repetitive tasks. The event, scheduled for February 17, 2026, aims to teach participants how to create modern ingestion systems where AI facilitates mechanical work, allowing engineers to focus on validation and safety. The workshop will include hands-on activities such as building a pipeline for the GitHub API, using tools like DuckDB, and validating results with a dashboard. It will also discuss the importance of skepticism towards LLMs, advocating for governance through code with built-in validation and observability. Participants are encouraged to bring their own challenging APIs for a hackathon and to join the dlt community to further explore the potential of AI in data engineering.

In the blog post "The Last Mile is Solved by Slop," Adrian Brudaru, Co-Founder and CDO, discusses an innovative approach to simplifying the data modeling process in business intelligence (BI) applications by using a prototype workflow called "slop." This method leverages a Local Language Model (LLM) wrapped in a scaffold to automate the traditionally manual task of creating a star schema from raw data, significantly reducing the time required from days to minutes. The process involves several steps, including schema inference, question formulation, and model validation, using a mix of code, templates, and prompts that guide the LLM through a disciplined sequence of actions. This approach democratizes data engineering skills, allowing less experienced users to effectively perform tasks that traditionally required senior-level expertise. By automating the transformation of raw data into a structured format ready for BI tools, the prototype showcases the potential for LLMs to redefine the role of data engineers, shifting the focus from technical implementation to strategic decision-making and outcome definition.

Adrian Brudaru describes a new approach for automating complex data modeling processes using a prototype called "slop," which leverages large language models (LLMs) to transform raw data into a dashboard-ready star schema with minimal human intervention. The traditional manual slog of inspecting data, designing layers, and writing SQL is replaced by an automated workflow that bridges the gap between raw data and usable business insights. This is achieved through a scaffold that runs a pipeline, extracts schemas, and guides the LLM to build a star schema, reducing the time required from days to minutes. The process involves anchoring the data with evidence, defining goals through specific questions, and employing a scaffold as a programmable skill to guide the LLM. The prototype demonstrates that with the right structure, an LLM can automate complex tasks like dimensional modeling, commoditizing traditional data engineering skills and shifting the value to defining desired outcomes. The ultimate vision is for the system to achieve full autonomy, where human validation is replaced by automated testing, allowing the agent to self-correct and complete tasks without manual prompts.

Aman Gupta, a Data Engineer, explores the performance benefits of using Apache Arrow and ADBC over SQLAlchemy for EL pipelines that transfer data from DuckDB to MySQL. The experiment demonstrates a significant 3.7x speedup when adopting Arrow's columnar data format and ADBC for bulk loading, reducing the time from 344 seconds to 92 seconds. This efficiency is achieved by minimizing Python object handling and serialization costs, thereby shifting bottlenecks away from the CPU. Arrow's in-memory columnar format streamlines data movement, reduces compute costs, and enhances throughput by eliminating the overhead associated with row-based data structures. The use of dlt with Arrow further simplifies the pipeline architecture, ensuring fewer moving parts and easier maintenance while maintaining high performance.