Ask in plain words
Type the question the way you would ask an analyst, including nicknames, ratios and relative dates.
Ask your data anything. Verify every answer.
AIDA, your Artificial Intelligence Data Analyst, uses a language model to understand what you mean and deterministic code to decide what runs. Every phrase is traced, every calculation is exact, and your rows never leave the database.
0database rows sent to the model
5pipeline stages, each one inspectable
100%of answers show their SQL and lineage
Guard and budget
Prompt Guard 2 classifies the text as a normal question
Length, characters and your per-minute question budget checked
Only approved catalog labels are attached to the request
How AIDA works
Language understands. Code verifies.
Most text-to-SQL tools let a model write the query. AIDA splits the job: the model only interprets, and a deterministic compiler builds the SQL from definitions your team approved.
The LLM resolves names
It maps each phrase to an approved measure, grouping, value or time period, or asks you to choose when a name is ambiguous.
Code checks every link
Grounding, coverage, approved values and capability limits are verified before anything runs.
Exact, traceable answers
Read-only SQL runs inside AIDA. Charts, tables, SQL and lineage show exactly how the answer was produced.
Security by design
Six layers between a question and your data.
Pick a layer to see a real attack it is tested against. Each scenario is covered by an automated test in the AIDA backend suite.
Accounts and sessions
Passwords are hashed with scrypt. Sessions live in HttpOnly, SameSite=Strict cookies, are stored only as hashes, rotate at sign-in and expire when idle.
Attack scenario
Twelve password guesses against one account in a minuteThe account locks for 15 minutes after 5 failures, and the error never reveals whether the email exists.
Privacy you can inspect
What the model sees, and what it never will.
The interpretation model works from a compact catalog of approved labels. Everything that identifies your data or your people stays inside AIDA.
Your question text
Approved measure and grouping labels
Plain-language definitions you approved
Owner-approved filter values
The dataset’s date range and as-of date
What each source can and cannot answer
Capabilities
An analyst’s toolkit, without an analyst’s backlog.
Name resolution that asks
An LLM maps everyday words to your approved definitions. When a name is ambiguous or unknown, AIDA asks you to choose instead of guessing.
Calculations in code
Ratios, differences, share_of_total, running_total and percent_change are computed after aggregation, never estimated by the model.
Time that makes sense
Calendar periods, “last month”, trailing windows and explicit ranges resolve against the dataset’s as-of date. Forecasts are refused, not invented.
Joins you approved
Relational sources follow approved relationships for related records, archived history and above-average comparisons, with lineage for every table touched.
Interactive results
Switch between bar, line, area, donut and scatter charts. Search, filter, sort and page through tables, then export exactly what you filtered.
Dashboards without re-asking
Saved analyses store the validated plan, so refreshing a dashboard runs SQL directly with zero model calls.
Benchmarks
Measured, not claimed.
Real questions through the real pipeline, scored against independently written SQL. These figures are read directly from the recorded benchmark runs.
48.1% → 70.4%correct on the 27 questions every compared run answered: gpt-oss-120b · Previous pipeline → Qwen3.8 27B · AIDA 4 · prompt 1 · repair 0
3 → 0wrong or unsafe answers on those same questions
112/112answers with a correct plan that also returned the correct rows, across 5 runs
10/10requests correctly refused or clarified by Qwen3.8 27B · AIDA 4 · prompt 1 · repair 0
These questions are a regression set, not a blind test. See every run, chart and question →
Onboarding
From sign-up to first answer in about two minutes.
1
Create your account
The first account becomes the workspace owner and can review security events.
2
Tell us about your team
Company, role and the areas you care about shape your starting point.
3
Choose your data
Explore the demos, add a private logistics sample, or upload SQLite and approve its fields.
4
Ask and verify
Review what is shared with the model, then ask. Every answer shows its plan, SQL and lineage.
Give your team answers they can check.
Start with the demo sources, or bring a SQLite snapshot of your own.