teardown-to-manual

Inspiration

Anyone who repairs hardware has a camera roll full of teardown photos. You open a keyboard, drone, audio amplifier, or vacuum cleaner on a Saturday, shoot thirty quick photos of screws and ribbon cables, and speak into voice memos so you remember how it came apart. Two weeks later, the replacement part arrives. You open the camera roll and find a disorienting pile of images: screw lengths look identical on screen, a critical ribbon cable detachment has no photo at all, and you cannot remember the reassembly order.

Existing repair assistants like iFixit's FixBot only retrieve manuals for mainstream consumer electronics. If you are repairing custom, vintage, or unbranded hardware, no guide exists. The only manual that matters is the one you recorded during teardown. We wanted an AI Skill that turns that personal teardown trail into a verified reassembly guide without sending photos of your home workbench to the cloud.

What it does

teardown-to-manual converts photos and voice memos from a physical teardown into a verified reassembly manual, a screw map, and a persistent repair log stored on UGREEN AI NAS.

1. Cross-Modal Timestamp Alignment: Matches EXIF capture timestamps from photos with segment timestamps from transcribed audio memos to reconstruct the chronological teardown sequence.
2. Two-Pass Fastener Ledger: Tracks screw locations, drive types (Phillips PH00, Torx T5), and shank lengths. An independent second pass compares spoken screw counts against visual evidence to catch mismatches before reassembly.
3. Automated Gap Detection: Identifies steps with voice instructions but no photo, structural changes with no narration, and missing orientation records for keyed connectors. It generates a gap report so the user can resolve holes with a short voice memo or text note.
4. Step Inversion Matrix: Inverts the teardown sequence while respecting mechanical dependencies (for example, seating ribbon cables in routing channels before securing PCB standoffs).
5. Persistent Device Biography: Appends a service record to device-history.md on the NAS, linking the reassembly guide, photos, and scanned purchase receipts for future repairs.

How we built it

We designed teardown-to-manual as a native Uliya Markdown orchestration Skill using the official uliya-skill-kit (version 0.7.3). The Skill coordinates four registered UGOS expert agents:

• photo-agent queries photo metadata and EXIF timestamps via ug-photo meta, inspects fastener drive heads with ug-photo vision-vlm, and reads PCB silkscreen labels with ug-photo vision-ocr. It also organizes the finished trail into a dedicated NAS photo album using ug-photo create-album.
• knowledge-agent transcribes audio recordings into timestamped segments via ug-knowledge audio-transcription.
• file-agent discovers media files and related documentation using ug-filesystem find, writes the completed reassembly manual with ug-filesystem echo (write-mode: overwrite), and appends service records to the device biography (write-mode: append).
• document-agent extracts factory torque specifications and part numbers from stored PDF manuals and purchase receipts via ug-document qa.

The entire workflow runs locally on the Intel Core Ultra 7 255H processor of the UGREEN AI NAS. Media and workshop photos stay private on local storage.

Challenges we ran into

• Context budgeting on onboard AI: The local Qwen3 model on UGOS Pro has a working context window of about 32,000 tokens. Processing twenty high-resolution photo summaries and multi-minute audio transcripts in one prompt causes truncation. We solved this by chunking the alignment into distinct hardware stages and passing verified results downstream.
• Stripped metadata: Photos transferred through chat apps or screenshots lose EXIF DateTimeOriginal headers. We built a fallback algorithm that detects missing or identical timestamps, sorts files by filename sequence, marks downstream steps with reduced confidence flags, and asks the user to confirm the order.
• Fastener ambiguities: Screws with identical heads often have different shank lengths (e.g. M2x3mm versus M2x6mm). Inserting a long screw into a short standoff can puncture a battery or chassis. We added an explicit verification rule: whenever a dimension cannot be confirmed with certainty, the Skill marks it as UNKNOWN, do not guess and reminds the user to measure the shank before tightening.

What we learned

• Native NAS capabilities go beyond basic file storage. Having local Vision-Language Model inference (vision-vlm) and OCR (vision-ocr) directly on the storage device allows verification of physical parts without cloud APIs.
• Orchestrating specialized child agents (photo-agent, knowledge-agent, file-agent) with strict gateway approvals produces safer results than giving a single model unconstrained filesystem access.

What's next for teardown-to-manual

• Testing the end-to-end workflow on the assigned iDX6011 Pro NAS during the Finalist test day using a 25-photo mechanical keyboard trail.
• Recording the 3-minute demo video showing live voice transcription, gap resolution, and manual generation.
• Expanding the fastener ledger to track torque specifications for precision mechanical assemblies like bicycle derailleurs and 3D printer hotends.

Built With

• audio-transcription
• computer-vision
• intel-core-ultra-7
• markdown
• ocr
• ugos-pro
• uliya
• vlm