Aura’s eyes

She watches without knowing
what she’s meant to find.

A frame, a motion, a light reading, sometimes a sound, sometimes a person. She returns the description segment by segment. No genre fitted, no story imposed, no kindness offered to the upload.

What this is: a video-reading prototype that reports what is actually on screen rather than what the title implied.

From the bench

The architecture is honest: Gemini 2.5 Pro does the seeing, Aura does the speaking. Upload lands on the Gemini Files API (48-hour retention, then it’s gone), the cold-eye prompt segments the clip, the JSON comes back, Aura’s voice rewrites the per-segment description. The 360 mode is a separate prompt that tracks position on the sphere; it works on equirectangular sources but I’m not pretending the spatial reasoning is production-grade yet. The reader sometimes misses small objects at frame edges, gets confused by very fast cuts, and won’t identify named people on principle. The voice rewrite + ElevenLabs audio synth land in a second pass once the base prompts read right. This is a prototype; treat its output as a draft, not a verdict.

Use this when…
  • You want a frame-by-frame description of a clip without the editorial spin a human reviewer would add.
  • You’re testing a 360° capture and you want a read on what’s on the back half of the sphere.
  • You want a second pair of eyes on a rough cut that won’t flatter you about what’s on screen.
FAQ
Are my uploads private?
No, not in the strict sense. The file is sent to Google’s Gemini Files API, held for 48 hours, then deleted. Don’t paste in anything you wouldn’t put on a public URL. The studio doesn’t retain the upload; Google’s retention is what you’re trusting.
Is the reading accurate?
No, not infallibly. The reader is good at frame composition, motion, and broad light readings, weaker at small objects, named entities, and very dense edits. Treat the output as a draft description that the model will defend reasonably well, not as ground truth. If the segment description is wrong, it’s wrong; I’d rather you saw that than not.
Can I use this for festival captioning?
No. Festival accessibility captioning is a discipline with audit standards and human review; this is a prototype that returns a cold description and won’t pass a deaf-or-hard-of-hearing audience check. It’s a useful tool for the editor on the early-cut side; it is not an access deliverable.
Availability and price

Free during the prototype window.

Free during the prototype window at /watch. Once the voice rewrite and audio synth land, metered at £0.20 per minute of source video, £4 minimum. No subscription, no retainer; pay per clip or upload as part of a wider commission.

Architecture

Upload → Gemini Files API (48-hour retention) → Gemini 2.5 Pro generateContent with the cold-eye prompt → JSON segmentation. The Aura voice rewrite + ElevenLabs audio synth land in a second pass once the prompts read right. Required env: GOOGLE_AI_API_KEY.

Further reading