Industry Scenario — Retail & E-commerce

The team: a merchandising + data team at a mid-size e-commerce company. The goal: on-brand product descriptions, a demand forecast that retrains itself, and edge suggestions on kiosks — all owned.

The journey

Build · 1 — connect your data

• Connect raw clickstream and sales prefixes to the catalog.
• Use a glob pattern (sales/2026_Q4_*.parquet) so the same config consumes new partitions automatically.

Build · 2 — explore the catalog

• Merchandising analysts search the catalog and use lineage to pick the feature table that's a direct product of the billing feed — not a stale copy.

Build · 3 — train the model

• Fine-tune a description model on thousands of past listings (attributes → description) in the instruction format. The model writes drafts in the brand voice.
• Post-train with GRPO where the reward is schema-valid JSON for product attributes — the model learns to emit exactly the fields downstream needs.
• Pipeline a weekly forecast retrain over the fresh sales glob with a GPU smoke check.

Ship — one base, many brands

• Deploy one base model and attach LoRA adapters per brand/region (US/EU/APAC listing styles). Tuning a new brand = train a small adapter + attach — no full redeploy.
• Canary a new adapter at low traffic before cutover.

Monitor — catch the season shift

• Drift schedule on the forecast model flags the autumn shift early, so the weekly retrain happens before accuracy visibly drops.

Own it — edge, no network needed

• Export a small fine-tune to MLX/ONNX for in-store kiosks — on-the-spot category suggestions with no round-trip.

What you own at the end

• On-brand description generation + self-retraining forecasts.
• Per-brand adapters on one endpoint (cheap to extend).
• Edge-deployable weights.

Key docs

Connect · Explore · Train · Fine-tune · Post-train · Adapters & canary · Export · Monitor