Tasks
Data, code and calculations
Arithmetic that is actually calculated, dates in Bikram Sambat, structured data pulled out of free text, and a sandbox for running code, SQL and a real browser.
Models are unreliable at exactly the things computers are reliable at. These tools exist so the answer to “what is the total?” is computed rather than predicted.
Where it runs: in AI staff runs, for staff given these tools.
Everyday accuracy
| It can | What that looks like |
|---|---|
| Calculate | Real arithmetic, not a model’s guess at a sum |
| Tell the time | The current date and time in both Gregorian and Bikram Sambat, on Nepal time by default |
| Convert a date | AD to BS and back |
| Extract structured data | Free text turned into named, typed fields you specify |
The finance-facing staff roles are written to push every number through the calculator rather than doing it in their head, for the obvious reason.
Date conversion is a real conversion against attested calendar data, not an approximation — which is why a staff member can be trusted with “what is the deadline in BS?” in a way it cannot be trusted with mental arithmetic.
Extraction is for parsing, not perception: give it a paragraph and a list of
fields — invoice_number, amount, due_date — and you get exactly those
fields back as data. It reads text, not scans.
The sandbox
Heavier work runs in an isolated sandbox, jailed to the task it belongs to.
| It can | What that looks like |
|---|---|
| Run code | Python or JavaScript, returning output, errors and any files produced |
| Run a shell command | Files, git, package installs, in the same sandbox |
| Work with files | Read, write, list and delete inside the task’s own directory |
| Query a database | Parameterised SQL against a connection an admin configured — read-only unless writes are explicitly allowed |
| Drive a browser | Navigate, click, type and extract on pages a plain fetch cannot reach |
The sandbox browser is the answer to pages that need JavaScript or a login; web search and reading covers the ordinary public web more cheaply.
What bounds it
Access level. Every tool is marked read, write, or destructive, and a staff member can only use tools at or below its own ceiling. A Read-only staff member can query and calculate but cannot write files or run a write query.
Configuration. Code execution, the shell, the browser and SQL all need infrastructure an administrator has set up — a sandbox runtime, a named database connection. Where that does not exist, the tool is simply not available, and the staff member will say so rather than pretending.
Isolation. The file workspace belongs to one task. Nothing written during one job is visible from another.
SQL connections are named and configured centrally. A staff member cannot point itself at an arbitrary database, and writes are off unless an administrator has deliberately allowed them.