How it works
From question to answer in five steps
No SQL, no dashboard building. If you can ask a colleague, you can ask AutoAnalyst.
- 1
Sign in — or use the demo
One click opens the live demo with credentials prefilled. Or create a free account in seconds.
- 2
Pick a dataset
The Olist e-commerce demo is selected for you. Or drop in your own CSV, XLSX or Parquet file.
- 3
Ask in plain words
Type a question in English or Hinglish — no SQL, no column names, no dashboards to build.
- 4
Watch the agent work
It plans, writes SQL, fixes its own errors and runs real statistics or ML — every step streamed live.
- 5
Get answers you can use
A chart, a cited answer and the exact SQL to copy or export. Then schedule it as an emailed report.
Under the hood
An agent pipeline you can watch
Every question runs through the same eleven steps of a LangGraph state graph. Each step streams to the chat as it happens — and the full trace is saved with timings, tokens and cost.
Understand
- Input guardinput_guard
Scores the question for prompt injection and unsafe requests.
- Routerrouter
Classifies intent: lookup, trend, comparison, forecast or “why”.
- Schema RAGschema_rag
Retrieves the relevant tables, columns, glossary terms and few-shot SQL.
- Plannerplanner
Breaks the question into up to 4 steps: main, trend, driver, supporting.
Query safely
- SQL gensql_gen
Writes DuckDB SQL for each step using your business definitions.
- Validatorsql_validate
Parses the SQL AST: SELECT-only, allowed tables, automatic LIMIT.
- Executor self-correctsexecute
Runs read-only with a timeout. On error it feeds the message back and retries (≤3).
Analyse & explain
- Analystanalyst
Mann-Kendall trends, bootstrap CIs, SHAP drivers, Prophet forecasts.
- Chartchart
Picks the right chart and emits a typed ChartSpec for the UI.
- Narratornarrator
Streams a short answer that cites only numbers from the results.
- Output guardoutput_guard
Masks PII and checks every number in the answer against the data.
Features
An analyst, not a chatbot
Everything you need to trust an AI answer about your numbers — and to keep getting it.
Real data science
Mann-Kendall trend tests, bootstrap 95% CIs, XGBoost + SHAP drivers and Prophet forecasts — not just a SELECT.
Mann-KendallSHAPProphetGuardrails everywhere
Prompt-injection scoring, SELECT-only SQL validation, read-only execution, PII masking and a hallucination check.
sqlglotread-onlySemantic layer
A business glossary (“revenue = price + freight on delivered orders”) and schema RAG keep metrics consistent.
glossaryRAGAuto-EDA on upload
Upload a file and get a profile instantly: types, missing values, distributions, correlations and outliers.
CSVXLSXParquetScheduled reports & alerts
Turn any question into a cron-scheduled HTML email, or alert when a metric crosses a threshold.
cronemailEval dashboard
120 gold questions measure execution accuracy, judge score, hallucinations, latency and cost — in CI too.
benchmarkCITrace viewer
A waterfall of every step with timings, tokens and cost, so you can see exactly how an answer was made.
spanscostModel-agnostic
Groq, Gemini, Claude, OpenAI or a local Ollama model — switch in Settings, with automatic fallbacks.
LiteLLMOllama
Try it
Questions people ask
Click one — you’ll sign in to the demo and land in the chat with the question ready to send.
- Which category's revenue fell in the last 6 months, and why?Root cause
- Forecast total revenue for the next 3 monthsForecast
- Pichhle 6 mahine mein sabse zyada revenue kis category ka tha?Hinglish
- Top 10 sellers by revenue in 2018Ranking
- Is average delivery time trending up? Is it significant?Trend test
- Compare review scores for late vs on-time deliveriesComparison
- Which states have unusual order volumes this year?Outliers
- What drives low review scores?Drivers
After you log in
A quick tour of the app
Everything lives in the left rail. Here is what each screen is for.
Chat
/chatWhere you ask questions and get answers.
- Streams each agent step, the SQL (with self-corrections) and the result table
- Chart, a cited answer and insight cards with p-values and confidence intervals
- Copy SQL, export CSV, pin charts and keep a history of conversations
Benchmark & trust
Accuracy you can check
120 gold questions — easy, medium, hard and Hinglish — run against the agent in CI on every pull request. A drop below the bar fails the build.
Result set matches the gold SQL’s result
A second model grades the written answer
Answers citing numbers not in the data
End-to-end time per question
LLM spend, from token counts
Guardrails on every question
Safety is enforced in code, not left to the prompt.
- Prompt-injection scoring on every question
- SQL parsed to an AST: single SELECT only
- No DDL / DML, ATTACH, COPY or file readers
- Table allow-list and automatic LIMIT
- Read-only DuckDB with query timeouts
- PII masking before answers are shown
- Every number in an answer checked against the results
FAQ
Questions, answered
Anything else? Open an issue on GitHub — or just ask the demo.
Which data does AutoAnalyst use?
Only the dataset you select. Out of the box that is the Olist e-commerce demo; after you upload a file, questions run against that file’s own isolated schema. The model sees table and column descriptions — never your whole dataset.
Is my data safe?
Queries run against a read-only DuckDB connection with timeouts. Every generated statement is parsed and rejected unless it is a single SELECT on allowed tables (no DDL, DML, ATTACH, COPY or file readers). Answers pass a PII-masking step before they reach the screen.
Which LLMs does it support?
Any model LiteLLM can call — Groq, Google Gemini, Anthropic Claude, OpenAI or a local Ollama model. Pick a model and add your own API key under Settings; a fallback chain kicks in if a provider fails.
Can I upload my own data?
Yes. Drag a CSV, Excel (XLSX) or Parquet file onto the Datasets page. Types are inferred, columns get LLM-written descriptions, the schema is indexed and an Auto-EDA report is generated — then you can chat with it.
What is the demo dataset?
A synthetic dataset with the same 9-table schema as the public Olist Brazilian e-commerce data, generated with a planted, known story (an electronics decline driven by late deliveries in RJ) so the demo and the benchmark have ground truth. The real Kaggle Olist dataset (~99k orders) can be added too.
Is it free and open source?
Yes — the code is MIT-licensed on GitHub. You can self-host it with a free Groq key, or run fully offline with a local Ollama model.
How accurate is it?
It is measured, not claimed: a 120-question benchmark (easy, medium, hard and Hinglish) compares result sets against gold SQL and runs in CI on every pull request. The latest numbers live on the Eval dashboard.
Built by
The person behind AutoAnalyst

Munna Kumar
@CodeWithMunnaXData Scientist & Full-Stack AI Engineer
“I build AI that ships — not just AI that demos.”
Data pipelines, ML models, GenAI agents, full-stack apps, production deployment — I own the whole journey myself, end to end. And I teach everything I build to 1000+ students on YouTube.
- Years experience
- 2+
- Projects shipped
- 10+
- Core domains
- 6
- Students taught
- 1000+
Your first answer is one click away.
Open the demo with the Olist dataset already loaded, or sign up and bring your own file.