Multi-Agent Data Scientist
Published:
A desktop application for building and running AI-powered data analysis workflows. The main screen is a node-based canvas where you wire together data sources, context, and agents; runs stream live thoughts, tool calls, and observations into an event log.
Node types
| Node | Purpose |
|---|---|
| Connector | Connects to a CSV folder or Athena warehouse |
| Context | Injects markdown documentation into the agent’s prompt |
| Metrics Extraction | Reads chart images / CSVs and extracts descriptions with a vision LLM |
| Recommender | Converts product hypotheses into a structured analysis plan |
| DS Agent | The data scientist — asks a question, reasons over data, streams thinking |
Recommender agent. Instead of asking a precise analytical question, you state a product hypothesis. The Recommender profiles the connected tables (row counts, date ranges, column stats, grain), detects experiment-assignment columns, estimates sample sizes and statistical power, and emits a structured AnalysisPlan — typed hypotheses, scored candidate analyses, agent assignment (DS vs. Senior DS), execution order, and guardrails. The DS Agent then executes the plan and produces a markdown report.
Stack
Tauri + React desktop front-end, FastAPI + SSE backend, Python multi-agent orchestration. Plugs into OpenAI, Anthropic, and POE for model access; AWS Athena and CSV warehouse connectors; DuckDB-backed warehouse tools; markdown retrieval system for in-domain context.
MLOps / engineering choices
Auto-saving canvas with named workflow snapshots, JSON export/import, validation banner that distinguishes errors (block run) from warnings (allow run), live event-log streaming with one-shot cancellation, end-to-end pipelines from data ingestion through model deployment and monitoring.