The key problem we're solving: Current AI analytics tools try to connect directly to data warehouses through semantic layers, leading to hallucination risks and reliability issues. Instead, we've built an architecture that lets data teams deploy focused AI agents that work exclusively with curated datasets. The AI agent also operates primarily on a text-to-Python basis within the confines of the curated datasets (DataFrames) you’ve created, meaning users can do much more than just simple data pulls and pivoting.
Here are some of the technical notes:
- Dataset-based architecture leveraging DataFrames: Semantic layers are optional. Point agents at specific datasets (Python DataFrames you’re created) for faster deployment and reliable results
- Python-first analysis: Goes beyond text-to-SQL with support for machine learning and stats models and complex analysis without the end-user having to write any code
- Kernel management system: Each user gets their own Python kernel that stays in sync with source data, ensuring consistent analysis without affecting other users. So a user can edit filters in Fabi.ai Smart Reports to update the DataFrame leveraged by the AI.
- Universal data connectivity: Connect to any data source and merge multiple sources in memory - from spreadsheets to data warehouses
- Built-in validation: Agents can validate their own work and check for edge cases & data quality issues with transparent access to underlying code
We built this because we've both been frustrated with some of the shortcomings of more general AI tooling when it comes to data use cases. After experimenting with various approaches, we found that specialized agents with clear boundaries worked better than a one-size-fits-all AI solution. We focused on building out the following infra:
- Autonomous AI agent system that handles its own package management and tools
- Dynamic memory management for maintaining context without overloading
- Real-time data synchronization between shared reports and individual user kernels
- Validation systems to ensure code runs successfully before returning results
We currently see folks most often use Analyst Agent to build dedicated data agents for marketing, product, and customer success analytics. These agents are also a way to automate routine data exploration and validation tasks, and help business teams explore data independently inside controlled guardrails set by the data team or person building this agent.
We’d love for you to check us out and provide feedback on the architecture, how the agents handle complex queries, edge cases you run into, or any other question/requests/criticisms. The agents are included in our free tier.
Of course, happy to answer any technical questions about how we built this or discuss our approach to AI-assisted data analysis! Thanks!
