The Practical AI Round

At these companies the practical round is usually a few hours of take-home or a 90-minute pairing session, and the brief is often deliberately thin: a chat interface, a summariser, a document Q&A, something with tool calls. Thin briefs are the test — see working in ambiguity.

What separates submissions, roughly in order of how often it's the deciding factor:

The non-happy states. Loading that isn't a spinner blocking the page, an error the user can act on, a retry, an empty state, a request the user can cancel mid-stream. Cancellation specifically is the detail that gets noticed, because it means you thought about a user changing their mind on a slow response — and because getting an AbortController correctly wired through a stream is a real skill.

Streaming done properly. Tokens rendering as they arrive without layout thrash, a visible stop control, partial output preserved when the stream breaks. If you only have time for one advanced thing, this is it.

Handling the model being wrong. Malformed JSON, a refusal, an empty response, a truncated tool call. A submission that degrades gracefully instead of throwing signals someone who has done this before. Never render model output as raw HTML — treat it as untrusted, because it is.

Restraint. Four hours spent on scope you chose is worse than three hours on a smaller thing done well plus a README explaining what you cut. Don't add auth, a database, and a design system to a summariser.

A key in the repo, ever. Server-side proxy, .env.example, and a note in the README. Committing a key is an instant fail at security-conscious companies and it happens constantly.

Ask up front whether AI tools are allowed — usually yes — and if so, use them and be ready to explain every line you submitted.

This round is the closest thing to the actual job at these companies, and it's weighted accordingly. It's also the one where preparation transfers directly: having built one streaming interface with cancellation before means you spend the session on judgement rather than mechanics.

• The happy path is table stakes; loading, error, empty, and cancel states are where the score is.
• Streaming with a working stop control is the single highest-value thing to get right.
• Wire cancellation through properly — abort the request, not just the UI.
• Handle malformed, refused, empty, and truncated model output without crashing.
• Never render model output as HTML; treat it as untrusted input.
• Never commit an API key — proxy server-side and ship a .env.example.
• Keep scope small and put what you cut, and why, in the README.
• Ask whether AI tools are permitted, and be able to explain every line you submit either way.