Product Information Management Platform

Overview

A Next.js + Prisma PIM (Product Information Management) system with a Digital Asset Manager (DAM) baked in. Multi-brand catalog support with brand-specific rules in JSON, per-channel readiness validation, AI enrichment that proposes but never overwrites, and a clean operator UI for the catalog team.

Problem

Multi-brand retailers like Mike Sport need a single product hub that respects brand-specific data rules (Adidas requires X attributes, Crocs requires Y), validates against per-channel requirements (Shopify, marketplace, retail), and uses AI for enrichment without giving up control of master data. Off-the-shelf PIMs were either rigid or too expensive for the size of the operation.

Approach

Build a governance-first PIM where brand rules and channel profiles are configuration, not code. Treat AI as a contributor with proposal rights, not commit rights. Never silent-overwrite anything: every machine-generated value goes to an EnrichmentTask queue with confidence and source attribution.

Architecture

• App: Next.js 15 App Router, TypeScript, Tailwind, shadcn/ui primitives.
• ORM: Prisma over SQLite by default (swap to Postgres for production deploys).
• Validation: Zod schemas for brand rules, channel profiles, and runtime input.
• Asset storage: SHA-256 deduplication; any backend (S3, R2, local) works behind the storage adapter.
• Brand rules: JSON-defined per-brand attribute requirements parsed by the rules engine on every product change.
• Channel profiles: Per-channel JSON specs that drive the validator's structured issue reports.

Tech stack

• Frontend / backend: Next.js 15, TypeScript, Tailwind, shadcn/ui
• ORM and DB: Prisma, SQLite (default) / PostgreSQL (prod)
• Validation: Zod
• Storage: Adapter pattern (S3 / R2 / local)

AI work

• Image intelligence stub: heuristic-based by default, swappable to OpenAI Vision, Google Vision, or AWS Rekognition behind a single interface.
• Enrichment LLM stub: proposes values to an EnrichmentTask queue with confidence and source tracking. Never overwrites.
• Completeness engine: weighted scoring across required attributes, locales, imagery, and SEO metadata, with weights tuned per brand.
• Channel validation: structured issue reports point operators directly at what's missing for which channel.

Engineering highlights

• Non-negotiable integrity rules: AI never overwrites; every proposal traces back to a source; assets dedupe by SHA-256; locales never collapse silently.
• 12+ admin pages: products, assets, brands, channels, validation, imports, exports, activity, settings, and more, all with the same operator-grade UX.
• Pluggable everywhere: vision provider, LLM provider, and storage backend all live behind narrow interfaces, making vendor swaps a config change.
• Audit trail: every enrichment, every overwrite-by-human, every channel publication is timestamped and attributable.

Outcome

Production-ready, used as the catalog backbone for downstream commerce and analytics. AI proposals reduce manual data work without ever compromising the master data layer.

Lessons

• Brand rules are configuration, not code. The moment you hard-code one brand's quirks you've created a maintenance trap.
• "Propose, never overwrite" is a five-word architecture decision that prevents an entire category of AI failure modes.
• Operator UIs deserve the same care as customer-facing ones. The catalog team is using this every day.