feat: vision classification pipeline — condition scoring, listing quality, fraud signals #21

New issue

Open

opened 2026-04-04 18:43:45 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-04 18:43:45 -07:00

Owner

Summary

Snipe currently runs basic vision tasks (photo analysis, serial number check). A proper classification pipeline using cf-core's LLMRouter + CFOrchClient would enable:

Condition scoring — structured assessment from listing photos (box condition, visible wear, completeness)
Listing quality signals — photo count/quality, description completeness, seller history indicators
Fraud/misrepresentation signals — stock photo detection, description/image mismatch, VERO risk
Category-aware classification — electronics vs. collectibles vs. antiques have different condition rubrics

Architecture

Build on top of circuitforge_core.documents (Dolphin-v2 ingestion pipeline) + circuitforge_core.llm.router.LLMRouter for structured output. The existing app/tasks/runner.py + tasks/scheduler.py shim is the right integration point.

Relationship to trust score

The existing eBay trust score is seller-reputation-based. This adds a listing-quality layer. Both feed into a combined confidence score shown to the user before they bid.

Phases

Structured condition scoring from photos (Dolphin-v2 via cf-docuvision)
Description/photo consistency check (LLMRouter structured output)
Fraud signal flagging (stock photo fingerprint, VERO keyword scan)
Combined confidence score UI

circuitforge_core.documents — Dolphin-v2 ingestion
circuitforge_core.llm.router — structured output
circuitforge_core.resources.CFOrchClient — VRAM allocation for vision tasks
cf-core#8 (cf-docuvision service, merged)
Existing vision tasks: app/tasks/runner.py

## Summary Snipe currently runs basic vision tasks (photo analysis, serial number check). A proper classification pipeline using cf-core's `LLMRouter` + `CFOrchClient` would enable: - **Condition scoring** — structured assessment from listing photos (box condition, visible wear, completeness) - **Listing quality signals** — photo count/quality, description completeness, seller history indicators - **Fraud/misrepresentation signals** — stock photo detection, description/image mismatch, VERO risk - **Category-aware classification** — electronics vs. collectibles vs. antiques have different condition rubrics ## Architecture Build on top of `circuitforge_core.documents` (Dolphin-v2 ingestion pipeline) + `circuitforge_core.llm.router.LLMRouter` for structured output. The existing `app/tasks/runner.py` + `tasks/scheduler.py` shim is the right integration point. ## Relationship to trust score The existing eBay trust score is seller-reputation-based. This adds a listing-quality layer. Both feed into a combined confidence score shown to the user before they bid. ## Phases 1. Structured condition scoring from photos (Dolphin-v2 via cf-docuvision) 2. Description/photo consistency check (LLMRouter structured output) 3. Fraud signal flagging (stock photo fingerprint, VERO keyword scan) 4. Combined confidence score UI ## Related - `circuitforge_core.documents` — Dolphin-v2 ingestion - `circuitforge_core.llm.router` — structured output - `circuitforge_core.resources.CFOrchClient` — VRAM allocation for vision tasks - cf-core#8 (cf-docuvision service, merged) - Existing vision tasks: `app/tasks/runner.py`