# Old Forge Technologies - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Andrei Trimbitas](https://cerebralvalley.ai/u/creativeheadz) - **GitHub:** https://github.com/creativeheadz/wegweiser-eyewitness - **Demo video:** https://www.youtube.com/watch?v=fSsAfWAcBeQ - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/109 Wegweiser Eyewitness: when the agent goes dark, the user's phone becomes the sensor. Every IT helpdesk knows the call: "What does the screen say?" "It went black. There's some tiny writing at the bottom." When a machine won't boot, the monitoring agent on it can't report anything, and the first hour of the ticket goes on questions. Eyewitness is a step in front of the helpdesk. When a ticket arrives by email, the portal or the Wegweiser help button, Muse Spark reads it against a checklist of what a technician needs. When something is missing that only a person standing at the machine can see, it holds the ticket and sends the customer a link with a short list of steps, written from fixed templates. On the phone, the customer photographs the screen, films the power light or records a voice note. Each capture is checked on the phone and again on the server before any model sees it, so a blurred photo gets an exact retake instruction at once. - Muse Spark reads the crash screen, including the small stop code line at the bottom of the current black Windows 11 screen, and reads device labels. - SAM 3.1 follows the blinking light through the clip, and a few lines of code count the flashes and decode them. Two amber and three white mean "memory not detected" on a Dell. - Muse Voice Transcribe turns a voice note in any of 25 languages into text, and Spark summarises it in English. The findings go on the ticket as an internal note, and Wegweiser's triage takes the ticket from there. Security: customer evidence is untrusted, so the model never writes to the customer. Everything the customer reads comes from fixed templates. Readings travel inside an untrusted fence, and text that tries to instruct the AI, such as a sticky note asking for the admin password, is flagged and treated as data. Uploads are checked by content, links are scoped and expire, every public route is rate limited, and media jobs are bounded and leave nothing behind. Integration: Eyewitness plugs into existing service desks. Zammad pushes each new ticket through a signed webhook. Any other desk or automation tool uses the Evidence API with signed callbacks. Any agent can use the MCP server, and a Muse Spark agent requesting evidence through it is part of the repo. Open weights, on your own hardware: some support companies can't let a customer's photos leave their own network. A tenant can switch Eyewitness to Muse Glimmer, Meta's open-weights model (27.9B, Apache 2.0), running on hardware the company owns behind any OpenAI-compatible server: Ollama, llama.cpp, vLLM or a partner host. Every read for that tenant then stays on its network, with no fallback to the cloud even when the local server is down. On a desktop with an 8 GB RTX 3050, Glimmer read the crash screen's stop code correctly in about a minute at no cost, and the staging end-to-end check of this mode passed. Tasks as data: every task a customer can be given is a small file a technician writes, with the words the customer sees, a line saying when to ask for it and, for a photo, what to look for and the fields to fill. Spark reads any photo task from its file, so a printed test page or a monitor showing No Signal needed no new code. Beeps are counted in code, since a model can't hear them. A new task is a new file and a restart. Built during the event: everything in the repo was built between 26 Oct and 4 Nov 2026. Wegweiser itself, the monitoring platform with its helpdesk connector and triage, existed before the event and stays private. The repo holds only the new Eyewitness code, and the video marks existing parts with an "EXISTING WEGWEISER" badge. Testing: 166 automated tests with no database and no paid calls, plus the full Wegweiser suite. Nine end-to-end scenarios passed on a staging server with live model calls, including a sticky note that tries to instruct the AI and a ticket read on the tenant's own hardware by Muse Glimmer. The newest runs the catalogue's tasks, from a printed test page to a beep recording. One ticket with an intake read, a screen photo, a light video and a voice note costs $0.04 to $0.06 of Meta API, depending on how many SAM calls the light video needs. ## More from Meta Global AI Developer Hackathon - [Kanopi Lens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/106) - [CO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/107) - [NeuroLab AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/108) - [AUTOKEREN](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/110) - [Proofline](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/111) - [Cengkudu 800 MDPL](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/112) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/109. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt