01 The project

An AI-driven pipeline that writes tech packs.

Brief → AI extraction → tech-pack assets · co-authored with Claude Code

A footwear tech pack is the spec a factory builds from — bills of materials, points of measure, construction, grading. This pipeline reads what a designer already makes (sketches, 3D renders, written notes) and auto-populates those documents, applying the brand's own material, trim, and fit rules, then validates the result before anything reaches a sample line.

Project
AI tech-pack generator
Multimodal input to factory-ready spec documents
Context
Speculative Brand Brief
Built for a client engagement — not a portfolio demo
Stack
Vision + language models
Structured extraction · rule-based validation · built via Claude Code
Walkthrough

The pipeline, running.

This is content supply chain fluency made literal — a brief turned into the spec a manufacturer can actually build from.

02 My role

Sole builder, brief to working pipeline.

I designed and built every layer: the multimodal ingestion, the logic that populates each spec section, the brand-standards layer, and the validation rules — composed through an AI build interface, with the judgment to know when its output was wrong and the error handling was missing.

03 The problem

Tech packs are the blueprint. Making them by hand is the bottleneck.

Every garment needs a detailed spec before it can be made, and producing those by hand is slow and error-prone — it gates the seasonal calendar. Worse, an error caught late doesn't just cost an edit; it costs another physical sample round, measured in weeks and dollars.

04 What I did

Turned a designer's inputs into a populated, checked spec.

Ingest

Reads what designers already make

Sketches, 3D renders, and written notes go in as-is. No new input format, no extra step bolted onto the team.

Populate

BOM, POM, construction, grading

The pipeline fills the load-bearing spec sections from those inputs, instead of a person re-typing them by hand.

Apply standards

The brand's rules, baked in

Materials, trims, and fit rules follow the brand's own taxonomy, so the output reads on-brand, not generic.

Validate

Catches errors before sampling

Missing measurements, incompatible fabrics, and regulatory gaps are flagged in real time — at the desk, not at the factory.

05 How it works

From a designer's inputs to a factory-ready spec.

The pipeline moves left to right: a designer's existing work goes in, the model reads it, the spec sections are populated against the brand's standards, and a validation pass gates the output before it ships downstream.

The AI-driven process
01 · Input
Sketch / render / text
Whatever the design team already produces.
02 · AI
Extraction
Reads design intent and detail from each input.
03 · Populate
Spec sections
BOM, POM, construction, and grading filled in.
04 · Standards
Brand rules
Materials, trims, and fit rules per the brand taxonomy.
05 · Validate
Compliance pass
Flags measurements, fabrics, and regulatory gaps.
06 · Output
Tech-pack assets
The factory-ready spec — the sample on GitHub.

The validation pass is the load-bearing step. It catches the errors that would otherwise surface at the sample line, where each one costs a round — the same place the savings actually live.

1 UPSTREAM & CONNECTORS PLANNED PLM · 3D CAD · sketch & render apps designer-supplied inputs 2 INGESTION & DATA LAYER PLANNED connector services · object storage metadata store: project / season / style IDs 3 AI EXTRACTION & UNDERSTANDING BUILT vision reads sketches & renders · NLP parses written specs → design intent, measurements, BOM/POM, construction, grading 4 RULES & STANDARDS ENGINE PARTIAL applies brand materials, trims & fit rules today configurable per-brand taxonomy + rule engine (planned) 5 VALIDATION & COMPLIANCE BUILT measurement / fit · fabric · regulatory validators real-time flags · feedback loop back to the designer 6 OUTPUT & DELIVERY PARTIAL tech-pack generator · versioned GitHub output (the sample) PLM / 3D export · notifications (planned) 7 PLATFORM & OPS PLANNED Orchestration / workflow Monitoring & logging Model management Security & access control Cross-cuts layers 1–6 Built · ran in production Partial · built, not yet generalized Planned · designed, not shipped
The reference architecture. Solid blocks shipped to produce the sample tech pack; partial blocks run today but aren't yet generalized across brands; dashed blocks are designed, not yet built. Layer 7 cross-cuts the whole pipeline.
Where it plugs in · Direction, not shipped

Built to sit inside a real content supply chain.

The pipeline runs standalone today. The design goal is for it to live where the work already happens, rather than as a separate tool the team has to remember to open:

  • Plug into existing PLM and 3D tools instead of replacing them
  • Configurable to each brand's taxonomy for materials, trims, and fit
  • Scales with product volume across a season's worth of styles
  • Shared surface for design, tech design, and sourcing to work from one source of truth
06 The result

Proof at the spec level, not a speedup number.

3
Input types it reads: sketch, 3D render, text
4
Spec sections it auto-populates: BOM, POM, construction, grading
Live
Validation that flags errors before sampling, not after
1
Sample tech pack you can open and inspect

These are what the pipeline does, not a percentage it saves — the savings live downstream, in the sample rounds it prevents.

Carlos M. Cruz
Creative Technologist · Technical Artist