Answers from your business data
Ask your data.Get a number you can defend.
Tattva computes every defined figure in code from the definitions your analysts approved, checks its own answer before you see it, and shows its working line by line.
Works withBigQueryGoogle SheetsSlack
Queries BigQuery where the data lives · answer tables paste into Sheets as plain numbers · takes questions in Slack through Manas.
What was our P&L for Jul – Sep 2026, down to CM2?
- Intent
- a lookup: a P&L statement
- Metrics
- mrp · discounts · net_revenue · cogs · fulfilment_cost · marketing_spend
- Period
- Jul – Sep 2026 → 2026-07-01 to 2026-09-30
01 · Question
You ask in plain English.
Tattva reads the intent, the metrics you name and the period. A missing input becomes a question to you, not a guess.
02The ontology
Not just a dashboard.A definition that holds.
Tattva’s model of your business: which metrics exist and exactly how each is calculated, what your columns mean, how tables join, and which conventions apply. Answers, trackers and Try it all compute through the same code, so they give the same number.
- 01
Metric document
Your analysts write how the team calculates each number, query included, and upload it as .md or .txt.
net_revenue.md - 02
Stored definition
Your query, kept whole, with inputs like @start_date. Checked against BigQuery (V1–V12) and accepted in Review. A build never changes it.
@start_date · @end_date - 03
Families and conventions
“Revenue” covers several metrics? Its family decides: show every variant, ask which one, or use a default and mention the others with their figures.
revenue → net_revenue - 04
The Trail
Every answer keeps every step and every attempt, including withdrawn ones. Open it to see the SQL, the inputs and where each came from.
Trail · 4
03Question types
Lookups, trends, comparisons, whys.Choose your question.
Tattva first works out what kind of question you asked. The answer’s tables then pick their own chart, in code, from the shape of the figures.
What was our P&L for Jul – Sep 2026?
A P&L statement is drawn as a waterfall: each subtotal is a full bar, each line in between moves the total up or down.
04Trust
Computed, then audited.Checked before you see it.
Figures computed in code
For a defined metric the model cannot change the figure; it can only explain it. Any other arithmetic goes through a calculate tool that runs in code.
runMetric → stored definition
A withdrawn attempt, shown
When the self-check rejects a draft, the answer is written again, and a quiet note after it says “An earlier draft was checked and revised before this answer was shown”. The Trail keeps the withdrawn attempt and the reason.
Withdrawn by the self-audit
Personal data, by clearance
Each member has a PII tier: None, Aggregated, Unsafe aggregates or Full. The default is None. Queries that read personal-data columns a member isn’t cleared for are refused before they run.
None · Aggregated · Unsafe aggregates · Full
An audit log for owners
Role changes, PII grants and revokes, source uploads, Review decisions and denied access are recorded with a hashed IP and kept indefinitely.
rbac.denied · pii grant · review
What is not checked: a number shown without working behind it, such as a bare table cell with no Compute entry or SQL, is not verified by the self-audit. The Trail shows which is which.
05Where it works
One set of definitions.Everywhere you ask.
Reasoning
Ask a question; the agent analyses, plans, executes and audits its own answer.
Answer · Trail
Trackers
Saved questions re-run hourly, daily or weekly, on the same metric definitions, delivered to Email, Slack or Teams.
Hourly · Daily · Weekly
Reports
A Reasoning answer laid out as slides, exported with Print to PDF or as Markdown. A report is a snapshot.
Slides · PDF · Markdown
Boards
A page of answers you pinned from Reasoning, each card a snapshot of the answer it came from.
Pin to board
Configuration document
A Word file describing how Tattva is set up for your project, with a change log against the last download.
.docx
Slack, via Manas
Ask in a DM, with /manas or by @mention. It runs exactly as in Reasoning: same projects, same definitions, same checks.
DM · /manas · @mention
What is Tattva?
Tattva answers questions about your business data. You ask in plain English and get an answer with the figures, the reasoning behind them, and a trail you can check line by line.
Which data can it read?
BigQuery tables, and uploaded files: CSV, TSV and Excel spreadsheets, plus Markdown and text documents. BigQuery is the only warehouse; PDF, Word and audio files aren’t supported.
Does it copy or change my data?
No. Tattva only reads. BigQuery tables are queried where they are and their rows aren’t copied, and nothing is written back to your sources.
How are the numbers checked?
Defined metrics are computed in code from the stored definition before the model writes. A self-check compares every figure with the computed one and sends a wrong draft back; whatever is still wrong is corrected in code, and the answer says so.
How do I check one myself?
Open the Trail. The Compute card shows each figure with its SQL, filters and inputs, and where each input came from.
Who sees what?
Everything in a project is visible to its members, and there are no public links. Roles run from Viewer (reads, can’t ask) to Owner, and by default you see no personal data.
Where does my question go?
The question, a description of the data and the query results are sent to AI providers, mainly Anthropic.
How long does an answer take?
Most answers take one to four minutes. The answer ends with how long it took, what it cost and how much data it read.