Tech • IA • Crypto
Building the Best Agentic Analytics Harness: Powered by Claude, Built with Claude Code
7/10TL;DR
Omni reports major productivity and product gains after integrating Claude models into its AI analytics platform, enabling faster development and more reliable data querying.
KEY POINTS
Rapid engineering acceleration
A team of roughly 25 engineers saw a sharp increase in development velocity after adopting Claude-based tools in early 2025. Internal code activity rose significantly, with even senior leadership continuing to contribute directly to production code. The shift followed an experimental phase that quickly turned into full adoption once newer models demonstrated consistent performance improvements.
Natural language to data queries
Omni’s platform allows users to “talk to their data” by converting plain-language questions into structured database queries. Claude is used to interpret user intent and translate it into a semantic query, which is then executed against enterprise data warehouses. The system bridges the gap between human language and complex SQL operations.
Semantic layer as core infrastructure
A central innovation is Omni’s semantic layer, which standardizes how data is defined and accessed across large organizations. It resolves ambiguities—such as differing meanings of “last quarter”—and organizes thousands of datasets into a usable structure. The layer also enforces permissions and ensures users only access authorized data.
Context improves AI accuracy
The system enhances model performance by embedding business-specific context directly alongside data definitions. Additional inputs such as AI-specific metadata, sample queries, and example values help guide the model toward correct interpretations. This localized context significantly improves answer quality compared to generic prompting.
From simple Q&A to agentic system
Early versions of the system handled single-question responses, but evolved into a full agentic architecture capable of multi-step reasoning. This includes task planning, iterative querying, and dynamic error correction. Introducing structured recovery mechanisms and better error messaging led to substantial gains in reliability on complex queries.
Model upgrades drive adoption
Transitioning from lighter models to more capable ones, such as Claude Sonnet, enabled deeper reasoning and longer interactions. Although this increased token usage, it unlocked more sophisticated use cases. Customers began using the system for queries that previously required hours of manual analysis.
Debugging through trace analysis
Detailed trace inspection became a key engineering practice. By analyzing how the AI system reasoned through failures, engineers identified structural issues, including inefficient multi-agent designs. Consolidating responsibilities into a unified “single-brain” architecture reduced unpredictable behavior and improved performance on complex tasks.
SQL generation breakthrough
A major improvement came from allowing Claude to generate native SQL instead of a constrained internal query format. By pairing this with a custom SQL parser, the system achieved more efficient, single-pass query execution. This reduced the need for multi-step query chaining and improved both speed and accuracy.
Interactive analytics and dashboards
The platform supports automatic creation of dashboards, visualizations, and exploratory workspaces. Users can refine AI-generated outputs through a graphical interface, aligning with a design philosophy of “AI for generation, UI for validation.” This hybrid approach improves usability and trust in results.
Continuous learning via feedback loops
The system incorporates user interactions back into its models and metadata, creating a continuous improvement loop. As business definitions and data evolve, the platform updates its understanding, maintaining accuracy over time.
Emphasis on evaluation and consistency
Internal evaluation systems measure output quality and consistency, addressing a key enterprise requirement: delivering the same correct answer repeatedly. Observability tools, especially execution traces, are treated as critical for diagnosing and improving system behavior.
CONCLUSION
Omni’s integration of Claude demonstrates how combining large language models with structured data layers and agentic design can transform analytics workflows, significantly increasing both development speed and the reliability of business insights.