Use cases / Finance
Spreadsheet analysis.
Python on your device.
Attach a workbook and ask a question. The model writes Python, and your browser runs it on the file with no network access, returning tables, charts and Excel files to download. The model reads the workbook’s text and the results, encrypted to its enclave.
An illustration of the workspace. The workbook and figures are made up, and nothing here is investment, tax or accounting advice.
The workflow
See it in the workspace.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
Encrypted before it leaves your device.
In your browser
- Runs the model’s Python on your file, with no network access
- Draws the tables and charts, and makes the files you download
- Encrypts the workbook’s text and Python’s output to the enclave
Undergrove’s relay
- Checks your sign-in and allowance
- Passes encrypted requests on, unable to read them
- Sees their size and time
Inside the enclave
- Turns the workbook into text
- Runs the model, which reads it and Python’s output and writes the code
- Tinfoil states nothing in it persists between restarts
Before sending, your browser checks the enclave’s hardware attestation against Tinfoil’s open-source release. Answers come back the same way, encrypted until they reach your browser. What verification proves
Features in the workspace today.
Python in your browser
numpy, pandas, matplotlib, scipy, scikit-learn and statsmodels, served from this site and run with the network closed.
Tables and downloads
show(df) returns a sortable table, and files saved to /out/ come back as downloads, up to 10 MB each.
Dashboards
Ask for a dashboard and every number is computed from the whole spreadsheet in your browser. It saves as one HTML file that works offline.
Kept variables
Each conversation keeps its variables between answers until its files change.
Web search
For outside context, the model can search the web; queries are screened for personal details first.
The limits, plainly.
- Check figures before relying on them. A model can misread a question or write code that runs but answers a different one; the code is shown above its results so you can read it.
- Nothing here is investment, tax or accounting advice.
- A large workbook’s text is cut to what the model can read, though Python still works on the whole file.
- Undergrove holds no certifications such as SOC 2 or ISO 27001, and hasn’t been independently assessed yet.
Try it with your own workbook.
Open the workspace, add your workbook and paste the request. Without live access, answers are prewritten samples.
Open the workspaceWhich cost centres grew fastest from Q2 to Q3? Show the top five as a table and a chart, flag outliers, and save the summary as an Excel file.