Scelo TUI¶
The three-pane Scelo workstation in a terminal. Drop a file in and the agent profiles it, cleans it to a fixed point, reads it, picks an analysis and runs it. You steer what it decided from chat. The whole session exports to Python, Jupyter, R, Excel and the Scelo IDE in one command.
Before you start
Scelo TUI installs from source rather than from a release, the analysis menu is eight entries against the IDE's thirty, and there is no session persistence. See Limits for the full list.
What you are looking at¶
The whole product is one screen. Left is the data, middle is what the agent decided, right is the result. Each pane has its own chat box at the bottom.
╭─────────────────────────────────────────────╮ │ SOFT · data ◂ focus │ │ │ │ file(s) │ │ └ book.csv │ │ 900 rows × 7 cols │ │ │ │ summary │ │ cleaned: 1 steps / 1 passes │ │ dropped: notes │ │ │ │ Rows represent one policy. │ │ Key analytical columns: │ │ - `inception`: Policy start date. │ │ - `sum_insured` and `premium`: Financial │ │ details of policies. │ │ Potential issues: │ │ - `claims_paid` has many missing values, │ │ which may limit analysis. │ │ │ │ ─────────────────────────────────────────── │ │ ask to change what the agent decided… │ │ │ │ │ │ │ │ │ │ │ │ ╭─────────────────────────────────────────╮ │ │ │ › ▌ │ │ │ ╰─────────────────────────────────────────╯ │ ╰─────────────────────────────────────────────╯
┌─────────────────────────────────────────────┐ │ TOOLS · models │ │ │ │ pipeline │ │ ✓ read file · 900 rows x 8 cols │ │ ✓ auto-clean · 1 step over 1 pass │ │ ✓ understand │ │ ✓ choose analysis · Value by segment │ │ ✓ run · `sum_insured` across `line` (5 │ │ segments) │ │ │ │ analyses │ │ ╔═════════════════════════╗ │ │ ║ book.csv ║ │ │ ║ 900 × 7 · 1 clean steps ║ │ │ ╚═════════════════════════╝ │ │ │ │ │ │ ┌────────────────────────────────┐ │ │ ├─▶│ ● Value by segment │ │ │ │ └────────────────────────────────┘ │ │ │ ┌────────────────────────────────┐ │ │ ├┄▶│ · Descriptive summary │ │ │ │ └────────────────────────────────┘ │ │ └┄▶ +5 more · /list │ │ To understand the distribution of premiums │ │ and claims paid by policy line and region. │ │ /run to change · /list for the menu │ │ ╭─────────────────────────────────────────╮ │ │ │ › … │ │ │ ╰─────────────────────────────────────────╯ │ └─────────────────────────────────────────────┘
┌──────────────────────────────────────────────┐ │ HARD · output │ │ │ │ table │ │ `sum_insured` across `line` (5 segments) │ │ │ │ segment n mean total share │ │ Motor 185 470,270 87.00M 21.7% │ │ Marine 184 455,146 83.75M 20.9% │ │ Liability 190 434,773 82.61M 20.6% │ │ Property 178 419,112 74.60M 18.6% │ │ Engineering 163 443,220 72.24M 18.1% │ │ │ │ plot │ │ sum_insured total by line · /charts │ │ Motor 87.0M │ │ Marine 83.7M │ │ Liability 82.6M │ │ Property 74.6M │ │ Engineering 72.2M │ │ │ │ flow │ │ ┌────────────────────┐ │ │ │ ● Value by segment │─┐ │ │ └────────────────────┘ │ │ │ ┌────────────────────┘ │ │ ▼ │ │ ╭──────────────────────────────────────────╮ │ │ │ › … │ │ │ ╰──────────────────────────────────────────╯ │ └──────────────────────────────────────────────┘
tab pane · drag to select+copy · ⏎ send (or paste a path) · ↑↓ history · /copy values · /help commands · ctrl-e export · ctrl-o model · ctrl-c quit
Every screen on this site is a real capture of the running program, not a mock-up.
The idea¶
Chat does not start work. Dropping a file does. The agent ingests, cleans, reads, picks and runs without being asked, and fills the panes as each stage lands. The bots exist to change what it decided, which is why every pane states what was done and why it was done.
That inverts the usual chat tool, where nothing happens until you compose a good enough prompt. Here the useful default has already run by the time you have finished reading the file name.
Sixty seconds¶
Pick a model, press Enter, and the panes fill. Then:
| you type | what happens |
|---|---|
/list |
the analyses that apply to this data, numbered |
/run 5 |
switch to that one, the HARD pane re-renders |
/charts |
every plot this data makes, full screen |
/export |
six files: cleaned CSV, pandas, Jupyter, R, Excel, .sce |
/live |
mirror the session into RStudio and Jupyter as it runs |
Nothing above needs a model. Slash commands are actions rather than conversation, and they keep working when the model is unreachable.
Where to go next¶
-
Two repos side by side,
bun install,bun link. No build step. -
The model picker, then dropping a file in.
-
What each one shows and why it says what it says.
-
Six artifacts, and how they are verified against the pane.