ADA Open-Source Analyst Computes Locally, LLM Optional
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- ADA runs the full pipeline locally with pandas — cleaning, type inference, schema detection, dashboarding, anomaly detection, and forecast — and exposes the calculation behind every chart and Ask ADA answer.
- The product works fully without an API key in deterministic mode; no model call is made for any question the rule-based parser can plan on its own.
- When an OpenAI key is configured, only two narrow model calls become available — a query planner and a strategic read — and the source states uploaded rows and cell values are never sent; only column schema and computed evidence reach the model.
- Model responses must match typed Pydantic schemas, per-request storage is disabled, and sessions are tagged with a hashed anonymous identifier for safety controls.
- The default model for routine strategic reads is gpt-5.6-luna (low reasoning), with gpt-5.6-terra (medium reasoning) reserved for ambiguous decisions; model calls are button-triggered and cached per evidence payload to prevent accidental spend.
- The codebase uses pure analysis functions and dependency injection at the model boundary so the business engine stays testable without Streamlit, network access, or API credits.
- The repo at github.com/saineshnakra/automated-data-analyst ships with ruff linting and a unittest suite that GitHub Actions runs on every push and pull request.
Why it matters: By keeping raw rows on-device and sending only schema plus computed evidence to OpenAI, ADA lets non-technical operators get AI-assisted analysis on business files they cannot share with a third-party model. The visible calculation under each answer turns the tool into an auditor-friendly analyst instead of a chatbot that returns unsourced prose.
