./tile-adhesive-selector

{
  "slug": "tile-adhesive-selector",
  "phase": "cancelled",
  "gate": null,
  "created_at": "2026-04-17T18:34:25.395Z",
  "updated_at": "2026-04-24T07:12:00.000Z",
  "notes": "Cancelled 2026-04-24 — сняли с активных, heartbeat-алёрт отключён. Чтобы реанимировать — phase обратно в 1-research."
}

# tile-adhesive-selector

- 2026-04-17T18:34:25.395Z · SPEC drafted via chat intake

# Fine-tune project: Tile Adhesive Selector

## Task
Given a customer description of a repair job and a 800-SKU product catalog, pick the right ceramic adhesive SKU.

## Inference format
- **Input:** Natural language customer repair description (text string)
- **Output:** Single SKU identifier (string)
- **Example:**
  ---
  Input: "Need to glue 60x60 porcelain tiles on an underfloor heating system in a bathroom, budget €15/sqm"
  Expected: KER-H40-FLEX-25KG
  ---

## Production context
- **Caller:** Shopify storefront agent on mw.ainmid.com, /api/recommend endpoint
- **Latency budget:** 2 seconds p95
- **Baseline in use:** gpt-4o-mini with RAG over SKU catalog (72% accuracy)

## Success criterion
90% top-1 SKU accuracy on 100-example hand-labeled validation set.

## Data
- **Labeled examples available:** 800
- **Source:** MasterWorks Postgres
- **Sample export possible:** Yes (CSV)

## Constraints
- **Privacy:** Public cloud OK
- **Hardware:** Apple Silicon M-series
- **License:** Permissive only (MIT/Apache)

• ./SPEC.md
• ./phase.json
• ./progress.md
• ./recommendation.md
• ./research-datasets.md
• ./research-models.md
• ./train.py

/Users/alphamachine/Projects/finetune-lab/projects/tile-adhesive-selector