Vision-based photo condition assessment (paid tier) #11

New issue

Closed

opened 2026-03-26 23:32:18 -07:00 by pyr0ball · 1 comment

pyr0ball commented

2026-03-26 23:32:18 -07:00

Owner

Description

Use a visual model to analyze listing photos for condition signals — scratches, dents, missing parts, screen damage, etc. — beyond what title keyword scanning can catch.

Approach

PhotoScorer stub already exists in app/trust/photo.py
Free tier: perceptual hash dedup only (already implemented)
Paid tier: vision model analysis via circuitforge-core LLM router
Local: moondream2 (fast, runs on CPU)
Cloud: Claude vision or equivalent

Signals to extract

Cosmetic damage visible in photos (scratches, dents, cracks)
Missing accessories or components
Screen condition (for laptops/phones)
Packaging condition

Privacy note

Photos are public eBay URLs — no user data involved
Results stored in photo_analysis_json on TrustScore (field already exists in model)

Acceptance criteria

PhotoScorer.analyze() calls vision model via LLM router
Results surfaced as additional red flags or score modifier
Gated behind paid tier check

## Description Use a visual model to analyze listing photos for condition signals — scratches, dents, missing parts, screen damage, etc. — beyond what title keyword scanning can catch. ## Approach - `PhotoScorer` stub already exists in `app/trust/photo.py` - Free tier: perceptual hash dedup only (already implemented) - Paid tier: vision model analysis via `circuitforge-core` LLM router - Local: moondream2 (fast, runs on CPU) - Cloud: Claude vision or equivalent ## Signals to extract - Cosmetic damage visible in photos (scratches, dents, cracks) - Missing accessories or components - Screen condition (for laptops/phones) - Packaging condition ## Privacy note - Photos are public eBay URLs — no user data involved - Results stored in `photo_analysis_json` on `TrustScore` (field already exists in model) ## Acceptance criteria - [ ] `PhotoScorer.analyze()` calls vision model via LLM router - [ ] Results surfaced as additional red flags or score modifier - [ ] Gated behind paid tier check

pyr0ball commented

2026-04-04 22:48:51 -07:00

Author

Owner

Infrastructure status (2026-04-04):

✅ trust_photo_analysis background task type registered in scheduler (PR #14)
✅ trust_scores.photo_analysis_json column exists in DB model
✅ tasks/runner.py downloads photo + calls LLMRouter stub for vision analysis
❌ PhotoScorer.analyze() not implemented — photo.py is phash-only; no vision model call yet
❌ Results not surfaced as red flags or score modifier

Superseded by #21 which is the full classification pipeline. #11 acceptance criteria (damage detection from photos) become Phase 1 of #21. Closing this in favour of #21.

**Infrastructure status (2026-04-04):** - ✅ `trust_photo_analysis` background task type registered in scheduler (PR #14) - ✅ `trust_scores.photo_analysis_json` column exists in DB model - ✅ `tasks/runner.py` downloads photo + calls LLMRouter stub for vision analysis - ❌ `PhotoScorer.analyze()` not implemented — `photo.py` is phash-only; no vision model call yet - ❌ Results not surfaced as red flags or score modifier **Superseded by #21** which is the full classification pipeline. #11 acceptance criteria (damage detection from photos) become Phase 1 of #21. Closing this in favour of #21.