ADA Turns Spreadsheets Into Auditable BI

ADA turns a CSV or Excel workbook into an interactive business-intelligence dashboard without requiring a configured BI platform—and its optional AI layer never receives uploaded rows or cell values.

The open-source project accepts CSV, XLSX, and XLSM files, cleans and profiles them, infers a business schema, and produces a Plotly dashboard. It also identifies anomalous periods, generates a guarded baseline forecast, explains material changes, and recommends the next investigation.

Deterministic analysis and optional AI

ADA keeps its trust boundaries visible across four layers:

• Calculation: Detects trends, drivers, anomalies, concentration, relationships, exceptions, and data-quality issues with deterministic, traceable logic.
• Conversation: Converts plain-English questions into typed pandas query plans executed locally. Each answer exposes its calculation.
• Interpretation: Turns computed results into prioritized investigations, clearly separating observations from causal claims.
• Optional AI: Uses model calls to plan queries that the rules cannot parse and to produce a strategic reading of computed evidence.

The deterministic engine remains authoritative even when a model is configured. The AI query planner receives only column names, types, roles, and the user’s question. The strategic-read feature receives calculated schema, summaries, evidence cards, recommendations, and user-supplied context. Uploaded rows and cell values are not included in either prompt.

Recommended reading

Harness Engineering Puts the Repository to Work

Responses must match typed Pydantic schemas, storage is disabled for the request, and a hashed anonymous session identifier supports safety controls. Calls are button-triggered and cached per evidence payload. The default model is gpt-5.6-luna with low reasoning; gpt-5.6-terra with medium reasoning is available for ambiguous decisions.

From spreadsheet to business decision

ADA automatically looks for:

• A primary outcome such as revenue, sales, profit, cost, amount, or units
• A time field for movement analysis, anomaly detection, and forecasting
• A useful segment such as product, category, channel, region, customer, or status
• Identifiers, missing values, outliers, concentration, and numeric relationships

Users can override the detected metric, date, and segment when a source schema is unusual. They can also focus the analysis on one segment and regroup it by the next useful dimension.

Core capabilities include a worksheet picker for multi-sheet workbooks, conservative cleanup and duplicate removal, a visible cleaning audit, an executive headline, four business KPIs, a downloadable Markdown brief, and a cleaned CSV export. Its anomaly radar flags periods outside a robust trendline band in the chart, evidence ledger, and recommended actions. The forecast includes month-of-year seasonality, an uncertainty band, and its backtested error beside the chart.

A movement waterfall reconciles the latest change by segment, while a segment-by-period heatmap shows intensity. The responsive Streamlit interface also includes a synthetic demo, file and row limits for predictable hosted performance, and plain-English Ask ADA queries covering totals, rankings, breakdowns, trends, growth, counts, and time or segment filters.

Local setup and privacy

ADA runs without an API key. To launch it locally:

''bash git clone https://github.com/saineshnakra/automated-data-analyst.git cd automated-data-analyst python -m venv .venv source .venv/bin/activate # Windows: .venv\\Scripts\\activate python -m pip install -r requirements.txt streamlit run app.py ''

Users may enter an API key in the session-only sidebar field. Private deployments can set OPENAI_API_KEY through the environment or .streamlit/secrets.toml, which should never be committed. The project warns against using an owner-funded key on a public deployment without authentication and spending controls.

The codebase separates Streamlit orchestration from the analysis engine, using pure functions and dependency injection at the model boundary. Tests cover unit behavior, privacy contracts, the pipeline, business logic, and rendering. GitHub Actions runs linting, the full test suite, and bytecode compilation on every push and pull request.

ADA is self-hostable on any Python-capable host running Streamlit and is released under the MIT license. Contributions are sought for deterministic metrics, natural-language question shapes, schema-detection fixtures, chart accessibility, additional file formats, and adversarial test datasets.