Turn CSV and Excel files into interactive BI dashboards

ADA is an open-source automated data analyst for operators who need answers from a spreadsheet without configuring a BI tool. Upload a CSV, XLSX, or XLSM file and ADA cleans it, detects its business schema, builds an interactive Plotly dashboard, flags anomalous periods, projects a guarded baseline forecast, explains material changes, recommends what to investigate next, and answers plain-English questions about the data with the calculation shown behind every reply. It's designed for a simple goal analytics software often makes hard: even a first-time user should be able to upload a spreadsheet and understand what's happening in the business.

ADA keeps four layers explicit rather than collapsing everything into generated prose: a Calculation layer (deterministic, traceable detection of trends, drivers, anomalies, concentration, relationships, exceptions, and data quality issues), a Conversation layer (plain-English questions turned into auditable pandas query plans executed locally, so every answer shows its math), an Interpretation layer (calculations turned into prioritized investigations, clearly labeled and never presented as causal proof), and an Optional AI layer (plans queries the deterministic parser can't read, and writes a strategic read over already-computed evidence - opt-in, and raw uploaded rows are never sent to it). The deterministic product is fully authoritative whether or not a model is configured at all.

Schema detection automatically looks for a primary outcome metric (revenue, sales, profit, cost, amount, units), a time field (for period movement, anomaly detection, and forecasting), a useful segment (product, category, channel, region, customer, status), identifiers, missingness, outliers, concentration, and numeric relationships - and users can override the detected metric, date, or segment without rebuilding the dashboard, or drill the whole analysis into a single segment slice and watch it regroup by the next useful dimension. Product capabilities include an anomaly radar (periods outside a robust trendline band, flagged on the chart, in the evidence ledger, and in recommendations), a guarded baseline forecast with month-of-year seasonality, an uncertainty band, and its own backtested error printed alongside the chart, a movement waterfall reconciling the latest change by segment, a segment-by-period intensity heatmap, conservative cleanup with a visible cleaning audit, an executive headline with four business KPIs, and a downloadable Markdown executive brief plus cleaned CSV.

Use ADA when you have a business CSV or Excel file and want a decision-ready dashboard, anomaly detection, and a forecast without modeling data in a traditional BI tool or trusting an unverifiable chatbot summary of your spreadsheet - it works fully with zero configuration and no API key, computing everything locally with pandas. It's particularly suited to operators who want every number backed by a visible calculation, and who need the analysis to stay authoritative and auditable rather than dependent on an LLM's own reasoning.

It's not built for arbitrary open-ended chat about unstructured data, and it's not a replacement for a full BI platform if you need deep custom modeling across many linked tables - ADA works from one uploaded file at a time (with multi-sheet Excel worksheet selection) and centers on business-metric-shaped data (an outcome, a time dimension, and a segment). The optional AI layer only activates when the deterministic query parser can't read a question, or for a higher-level strategic narrative, and even then it only ever sees schema and computed evidence, never uploaded rows or cell values.

Input is a CSV (comma, semicolon, or tab delimited), XLSX, or XLSM file uploaded through the Streamlit interface, with a worksheet picker for multi-sheet workbooks. Run locally:

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

Output is an interactive dashboard (charts, anomaly markers, forecast band, movement waterfall, segment heatmap), an evidence ledger showing the calculation behind every displayed signal, prioritized recommendations linked to that evidence, plain-English answers to typed questions (with their pandas calculation shown), and a downloadable Markdown executive brief plus cleaned CSV. No secret is required to run it; supplying your own API key unlocks the optional AI layer described above.

ADA's core analysis runs entirely on pandas and Plotly with no external service dependency. Its optional AI layer calls OpenAI's Responses API for two narrow, typed purposes: a query planner (invoked only when the deterministic parser can't read a question; it receives the column schema and the question, returns a proposed query plan, and ADA executes that plan locally, refusing unresolvable plans rather than guessing) and a strategic read (receiving only the calculated schema, summaries, evidence cards, and recommendations - never raw rows). No specific model or provider is fixed by default; each AI call waits for your confirmation before it runs, so nothing is triggered or spent without review. It deploys as a standard Streamlit app to any Python host.

Operators, analysts, and small teams who need answers from a business spreadsheet without configuring a BI tool or trusting an opaque chatbot summary, and developers who want a working, tested example of a deterministic-first analytics engine with a narrowly-scoped, privacy-conscious optional AI layer that never sends raw data to a model. It is licensed under MIT.