MAIA LiveCut watches the live bulletin, flags the moments worth a look, cuts them to the frame, and reframes them for every platform. Then it stops — and waits for an editor to say yes.
…kerajaan akan memastikan bantuan sampai kepada golongan sasar…
By the time a clip from the 8pm bulletin reaches social, the conversation has moved on. The bottleneck isn't editorial judgement — that part is fast. It's the mechanical work in between: finding the moment again, exporting it, recutting it vertical, retyping the caption twice. MAIA LiveCut removes the mechanical work and leaves the judgement exactly where it belongs.
Everything heavy runs in the cloud. The one stage that never gets automated is the one the newsroom actually cares about.
Built to answer the question a grant reviewer actually asks: who approved this, when, and what did it say at the time?
Every clip, title and caption lands in a review queue. Nothing reaches a platform without a named person approving it. The pipeline is designed so there is no path around this.
SuperAdmin, Socmed and Journo have genuinely different powers — enforced on the server, not just hidden in the interface. A journalist cannot approve their own clip.
Every edit, approval, rejection and publish is recorded with actor, role, timestamp and the before/after of what changed. Exportable as evidence.
An AI feature described vaguely is one nobody can govern. Here is each capability, what signal it uses, and what it explicitly does not know.
| Capability | How it decides | What it does not do |
|---|---|---|
| Suggested moments | Shot-change detection and speech/silence boundaries in the live buffer. | It does not understand what was said. It flags where something changed — a journalist decides if it matters. |
| Subject-tracked reframing | Per-column motion and detail analysis picks where the subject sits before cropping to vertical. | It is not face recognition. It identifies no one and stores no biometric data. |
| Transcript | Speech recognition with word-level timing, synced to the timeline. | It does not judge newsworthiness, and it will make errors on names and code-switching. |
| Drafted copy | A language model summarises the transcript into BM and EN captions. | It is instructed to report only what the transcript states. With no transcript it produces a blank placeholder rather than an invented headline. |
| Publishing | Nothing. There is no model in this step. | No clip reaches any platform without a named person approving it. This is enforced on the server, and logged. |
Every AI output is a draft. A person edits it, approves it, and is named in the log for having done so.
Live streaming, recording, transcription, smart reframing and transcoding run on Tencent Cloud managed services. MAIA builds the orchestration, the console and the governance.
The console runs against the real AWANI livestream. Pick a moment, mark it, and see the vertical clip come back.
Open the console