Shared AI services

Search over documents and video transcripts

• Semantic chunking: text is split where the meaning between sentences changes, with fixed-size chunks as a fallback.
• Hybrid search: vector search plus keyword boosting.
• Timestamps kept: transcript chunks keep their WebVTT timing and business metadata, so a search result can jump straight to the right moment in a video.

Coaching that never goes off-brand

Salespeople record videos, and the system gives them feedback:

• Classify first: a multimodal model watches the video and decides what kind it is, such as a walkaround, a thank-you, or a silent video.
• Choose, don’t write: the model must pick feedback word for word from copy that people wrote and approved. It never writes its own text, so the feedback can’t drift in tone or invent advice.
• A review screen: staff run the analysis in batches and check the results before sharing them.

One voice across products

• Shared pronunciation: personalised avatar videos use the same pronunciation layer as the vehicle-video pipeline.
• Lip-sync to our audio: speech is generated once, and the avatar lip-syncs to that audio.
• Fallback with an alarm: if speech generation misses its time budget, a cloned voice takes over, and an alarm fires if that fallback happens too often.
• No timeouts: the whole request runs under one deadline that is shorter than the API gateway’s timeout.