Delays by carrier
Median arrival delay in minutes, all 2025 flights.
Ai8 is a Jupyter-compatible workspace with an AI agent built in. Ask for an analysis and the agent reads your project, writes and edits cells, and runs them on your kernel — asking before each change.
Median arrival delay in minutes, all 2025 flights.
F9 and B6 have the worst median arrival delay, at 14 and 11 minutes. The chart is in cell 3, and its output is saved under the cell.
Explore, model and report in one place, with an agent that does the legwork you would otherwise type.
Projects, versions and environments in common — and assistant, LLM API and compute spend that stays inside an allowance.
Assignments, timed assessments and grading in the same workspace, with the assistant on or off per assignment.
The agent
Most notebook assistants hand you code to paste. The Ai8 agent uses the same files, notebook and kernel you do — and everything it does is written into the conversation, where you can open it.
It lists, searches and reads your notebooks, scripts and data — and PDFs, Word, PowerPoint and Excel files too. Each read is a line you can expand.
A proposed change arrives as a diff of that cell. Approve it and it lands in the live notebook, the way a collaborator's edit would.
It runs cells on your notebook's own kernel, and their outputs are saved under the cell. Quick checks run on a scratch kernel that touches nothing.
Longer work gets a checklist you can watch. What you ask it to remember — your conventions, a project's quirks — carries into the next conversation.
Control
An agent that can run code needs limits you can see. Ai8 puts them in the conversation, not in a policy document.
A new conversation asks before any edit or run. Allow one change, edits to one file, or every file — and revoke at any time.
Switch a conversation to Auto and tools run without asking. Access rules, allowances and the sandbox still apply.
Starting a runtime asks first, and names the machine and what it costs per hour. Nothing else the agent does starts one.
The agent reaches exactly what a cell of yours can reach — same container, same limits, same restricted internet.
A file or web page that tells the agent what to do is treated as content to read, not as a command to follow.
Ask before edits and actions unless you have allowed them in this chat.
The workspace
Sign in and your project is there: notebooks, files, a terminal, and a runtime the moment you need one. Nothing to install, nothing to keep in sync.
Notebooks, files and saved outputs are readable as soon as you sign in. A runtime starts only when something needs to run.
Standard Jupyter kernels in environments your team defines, with the packages it needs already installed.
Notebooks stay ordinary .ipynb files with their outputs. Download any of them, any time, and open them anywhere.
A terminal in the same runtime, and viewers for CSV, Excel, PDF, Markdown and HTML — next to your notebooks, in tabs you can split.
Inline code suggestions in every cell. Tab to accept, keep typing to ignore.
Automatic snapshots, named versions, cell-by-cell compare, and a restore that tells you what it will change first.
Editing carries on when the network drops and syncs when you are back. The interface says which state you are in.
See the dataframes and values in a Python kernel's memory without printing them.
For teams
A workspace for a whole group, run by its admins: who is in, what they can run, which models the agent and the LLM API use, and how much any one person can spend.
Owners, editors, commenters and viewers. Presence shows who is there; one person writes at a time, and hands over on request.
Plans set each member's assistant, LLM API and compute limits, on a weekly, monthly or custom reset, with per-person overrides. Members see a percentage, not a bill.
Admins publish environments from a package list — with a private PostgreSQL inside each runtime when a course or project needs one.
Admins pick the model the agent runs on from a catalog across the major vendors, and every call is metered to the member who made it.
Call any listed model from your own code — a small Python package or a plain HTTP endpoint — with no keys to manage: each runtime carries its own, and it stops when the runtime does. Included by the plan, and metered in credits against a limit of its own.
Python 3.12 · ds-base
For teaching
Ai8 grew up teaching data science to postgraduate cohorts. Courses, assignments and assessments are first-class — and so is deciding when the assistant helps.
Rosters, sections, announcements and materials. An assignment is an ordinary project, authored the way you author anything else.
Timed assessments start, end and submit on the server — even if a laptop closes or the network drops mid-exam.
On for the group project, off for the exam. While a timed exam is open, the assistant is off for that student everywhere.
Course staff can read the conversations students had on coursework. Students are told so in every brief — and their own projects stay private.
A grading queue, grades released when you decide, and a gradebook that exports to CSV for your VLE.
No webcam proctoring, no lockdown browser. Server-kept rules, version history and honest records instead.
Standard .ipynb with stable cell ids, and the standard kernel protocol. Your notebooks open in any Jupyter tool.
Notebook output renders on its own sandboxed origin, so code you run can never reach your session.
Download notebooks and files whenever you like. There is no proprietary format to export from.
Ai8 is available by invitation. Tell us about your team or course, and we will set up a workspace for you.