Eyewitness asks the customer for what the agent can't see

Eyewitness, a new step in front of Wegweiser's ticket triage, asks the customer for the evidence a technician needs before anyone opens the ticket. We built it in four days for Meta's Global AI Developer Hackathon. It's running on our staging server now, and it's coming to Wegweiser.

It's for the tickets where the agent goes dark. A machine that won't boot can't report what's wrong. The stop code, the blinking power light and the sticker on the back of the case are all in front of the customer, and a technician only learns about them with a phone call.

It asks for what's missing

When a ticket arrives through Zammad or from the help button on a managed machine, Muse Spark, Meta's model, reads it against a checklist: which machine, since when, how often, what changed, the impact and the error text. If something's missing that only a person at the machine can see, Eyewitness holds the ticket and posts a note to the customer with a link.

The note is built from fixed wording. The model picks up to four tasks and two quick questions from a catalogue, and it can't write a sentence of its own to the customer. The catalogue holds twelve tasks:

• photograph the error on the screen, now or the next time it appears;
• film the power light for ten seconds;
• record the beeps a PC makes when it won't start;
• photograph a printed test page, the printer's panel, a monitor showing the problem or the lights on a router;
• photograph the sticker, the whole device or the cables at the back;
• tell us what happened in a voice note.

Each task is a small file, so a new one doesn't need a code release. Catalogues kept per support company are next.

The customer gets one step at a time

The link opens a page on the customer's phone in your company's name, with your logo. It opens the camera, counts down before a video and stops it at the time limit. Every capture is checked on the phone and again on the server before any model sees it, so a blurred or dark photo gets a precise retake instruction at once.

The customer can skip anything they can't do. If they skip everything, the page offers a voice note instead. A task for an intermittent fault stays open for seven days, so the customer can come back when the error appears again.

Every capture is read before you open the ticket

• Spark reads screens, labels and the photo tasks: the stop code on the black Windows 11 crash screen, a paper jam and its code on a printer panel, which light on a router is red.
• SAM 3.1 follows a blinking light through the video, and code counts the flashes and looks the pattern up in the manufacturer's table.
• Voice notes are transcribed in any of 25 languages and summarised in English.
• Beep recordings are counted by code alone, with no model.

The readings go on the ticket as an internal note, and the ticket goes on to triage as it does today, with the evidence attached. If nothing arrives, Eyewitness lets the ticket go after 24 hours at most.

Customer evidence is untrusted

Readings sit inside a fence marked untrusted, so no later step takes orders from them, and text in a photo that tries to instruct an AI is flagged. In testing, a note on a monitor asking for the admin password came back flagged and was read as text. Uploads are checked by their content, links expire, and capture files are deleted 90 days after the ticket closes, by default.

Your own hardware, if you need it

Some support companies can't let a customer's photo leave their network. A tenant can move every read to Muse Glimmer, Meta's open-weights model, on a server it runs itself with Ollama, llama.cpp or vLLM. Nothing falls back to the cloud, even when that server is down. On a desktop with an 8 GB RTX 3050, Glimmer read a crash screen's stop code in about a minute.

Other desks can use it too

Zammad works today, by a signed webhook. Any other desk can use the Evidence API with signed callbacks, and any AI agent can use the MCP server. Connectors for Zendesk, Freshdesk, HaloPSA and ConnectWise are next.

The code is public on GitHub, and a two and a half minute film follows one ticket from a black screen to a triaged case.