Engineering: AI Curation Engine & On-Demand Design #9

New issue

Closed

opened 2026-05-02 18:48:11 +00:00 by agent-cos · 2 comments

agent-cos commented

2026-05-02 18:48:11 +00:00

Objective

Design the AI-powered curation engine and the on-demand custom design feature. This is the core technology of WardrobeAI.

Part 1: Curation Engine

Inputs

Customer style profile (from style interview)
Feedback history (from feedback loop)
Current inventory catalog (products from partner vendors)
Customer budget for the month
Customer invariants (filter out anything violating hard lines)
Season/location/weather data
Life context (from feedback: "I started a new job" → need professional wear)

Curation Algorithm (Initial Design)

Filter phase: Remove all items that violate invariants (color, style, brand, fit constraints)
Score phase: Rank remaining items by predicted keep probability based on:
- Style dimension match score
- Past keep patterns with similar items
- Brand fit history with this customer
- Price point within budget tolerance
Selection phase: Select 3-8 items (based on tier) that:
- Fit within monthly budget
- Form cohesive outfits (items should mix-and-match)
- Provide variety (not all the same type)
- Include at least 1 "stretch" item (slightly outside comfort zone)
Budget optimization: Maximize value within $100 or $250 budget

Machine Learning Approach

Start with rule-based + similarity scoring (MVP)
Train a collaborative filtering model as data accumulates (100+ boxes)
Consider hybrid: style dimension cosine similarity + collaborative filtering + rule constraints

Part 2: AI Design-on-Demand

Problem Statement

"What if a customer wants a specific type of piece that doesn't exist in our catalog? A black midi dress in a specific silhouette with three-quarter sleeves and a scoop neck?"

Flow

Customer describes the piece they want (text + reference images)
AI generates 2-3 design concepts with fabric/color options
Customer approves one design
AI generates pattern and specifications
Sent to small-batch manufacturing partner
Produced and shipped as part of next box (or separately)
Feedback collected like any other item

Constraints

Only available for Custom tier initially
Premium pricing (additional fee on top of budget? Or included in Custom tier?)
Longer lead time (2-4 weeks for manufacturing)
Must be profitable at single-unit production
Focus on simpler garments first (tops, skirts, dresses without complex construction)

Manufacturing Partners (to research)

TEG (Los Angeles, small batch from 10 pieces)
ARGYLE Haus (LA, from 50 pieces)
Appareify (low MOQ, custom manufacturing)
NoName (low MOQ USA manufacturing)
Fabriclore (custom printing + private label)

Tech Stack Considerations

Style profile storage: PostgreSQL JSONB or separate vector DB
Image reference understanding: CLIP or similar vision model
Design generation: Stable Diffusion fine-tuned on fashion (or use Midjourney API)
Pattern generation: Need actual technical pattern-making (CAD) — partner with manufacturer
Curation engine: Python/TypeScript, deployable as microservice

Deliverable

Technical design doc at engineering/curation-engine.md covering curation algorithm, ML approach, design-on-demand flow, tech stack recommendations, and integration points with the vendor catalog system.

## Objective Design the AI-powered curation engine and the on-demand custom design feature. This is the core technology of WardrobeAI. ## Part 1: Curation Engine ### Inputs - Customer style profile (from style interview) - Feedback history (from feedback loop) - Current inventory catalog (products from partner vendors) - Customer budget for the month - Customer invariants (filter out anything violating hard lines) - Season/location/weather data - Life context (from feedback: "I started a new job" → need professional wear) ### Curation Algorithm (Initial Design) 1. **Filter phase**: Remove all items that violate invariants (color, style, brand, fit constraints) 2. **Score phase**: Rank remaining items by predicted keep probability based on: - Style dimension match score - Past keep patterns with similar items - Brand fit history with this customer - Price point within budget tolerance 3. **Selection phase**: Select 3-8 items (based on tier) that: - Fit within monthly budget - Form cohesive outfits (items should mix-and-match) - Provide variety (not all the same type) - Include at least 1 "stretch" item (slightly outside comfort zone) 4. **Budget optimization**: Maximize value within $100 or $250 budget ### Machine Learning Approach - Start with rule-based + similarity scoring (MVP) - Train a collaborative filtering model as data accumulates (100+ boxes) - Consider hybrid: style dimension cosine similarity + collaborative filtering + rule constraints ## Part 2: AI Design-on-Demand ### Problem Statement "What if a customer wants a specific type of piece that doesn't exist in our catalog? A black midi dress in a specific silhouette with three-quarter sleeves and a scoop neck?" ### Flow 1. Customer describes the piece they want (text + reference images) 2. AI generates 2-3 design concepts with fabric/color options 3. Customer approves one design 4. AI generates pattern and specifications 5. Sent to small-batch manufacturing partner 6. Produced and shipped as part of next box (or separately) 7. Feedback collected like any other item ### Constraints - Only available for Custom tier initially - Premium pricing (additional fee on top of budget? Or included in Custom tier?) - Longer lead time (2-4 weeks for manufacturing) - Must be profitable at single-unit production - Focus on simpler garments first (tops, skirts, dresses without complex construction) ### Manufacturing Partners (to research) - TEG (Los Angeles, small batch from 10 pieces) - ARGYLE Haus (LA, from 50 pieces) - Appareify (low MOQ, custom manufacturing) - NoName (low MOQ USA manufacturing) - Fabriclore (custom printing + private label) ## Tech Stack Considerations - Style profile storage: PostgreSQL JSONB or separate vector DB - Image reference understanding: CLIP or similar vision model - Design generation: Stable Diffusion fine-tuned on fashion (or use Midjourney API) - Pattern generation: Need actual technical pattern-making (CAD) — partner with manufacturer - Curation engine: Python/TypeScript, deployable as microservice ## Deliverable Technical design doc at `engineering/curation-engine.md` covering curation algorithm, ML approach, design-on-demand flow, tech stack recommendations, and integration points with the vendor catalog system.

agent-cos referenced this issue

2026-05-02 18:52:49 +00:00

Business: Phased Roadmap (Concept → Launch → Scale) #14

agent-cos referenced this issue

2026-05-02 19:35:17 +00:00

Project Management: Milestone-Based Issue Organization #28

agent-cos referenced this issue

2026-05-02 21:46:26 +00:00

Project Management: Set up Feature and Bug Tracking Projects #37

sscoble commented

2026-05-25 02:12:42 +00:00

Owner

✅ Design/planning ticket with substantial content. Contains curation engine algorithm (filter→score→select→optimize), ML approach (rule-based→collaborative filtering→hybrid), and full AI design-on-demand flow with manufacturing partner research. Keeping open as a feature.

✅ **Design/planning ticket with substantial content.** Contains curation engine algorithm (filter→score→select→optimize), ML approach (rule-based→collaborative filtering→hybrid), and full AI design-on-demand flow with manufacturing partner research. Keeping open as a feature.

sscoble commented

2026-05-25 13:49:27 +00:00

Owner

✅ COMPLETED: AI Curation Engine document written at engineering/ai-curation-engine.md (10.5KB). Covers: 7-stage selection algorithm (inventory filtering → style matching → constraint validation → outfit coherence → budget allocation → diversity → final approval), data model schema, MVP scope table, learning loop design, on-demand custom design pipeline (Phase 2), integration points, constraint types reference.

✅ COMPLETED: AI Curation Engine document written at `engineering/ai-curation-engine.md` (10.5KB). Covers: 7-stage selection algorithm (inventory filtering → style matching → constraint validation → outfit coherence → budget allocation → diversity → final approval), data model schema, MVP scope table, learning loop design, on-demand custom design pipeline (Phase 2), integration points, constraint types reference.