This is an architecture case study — documentation of a production-grade system, not runnable software. Licensed under CC-BY-4.0. Based on a real enterprise brand production system. Client details are anonymized.
This repository documents the complete architecture of LoRA Studio — a SaaS platform that transforms brand guidelines into trained LoRA models for AI image generation with mathematically enforced visual consistency. The system orchestrates a 10-stage pipeline across 15 microservices to serve two personas: Marketing Creators (zero ML jargon) and Data Scientists (full parameter control).
The core technical innovation is the LOCKED/UNLOCKED Protocol — a deterministic method derived from the Flow Matching loss function that encodes brand identity directly into model weights, making brand drift structurally impossible rather than manually policed.
This case study is the architecture behind the concepts described in my LinkedIn articles:
- Como Criei um Sistema de IA que Transforma Diretrizes de Marca em 10.000 Assets (Business perspective — PT-BR, EN version publishing 2026-05-20)
- The Mathematics of Brand Memory: How Loss Functions Enforce Visual Identity (Technical perspective — publishing soon)
Every enterprise with a recognizable brand faces a silent tax: scaling visual consistency across channels without scaling headcount.
- A single major campaign costs $15K–40K when accounting for designer hours, review cycles, and rework from brand violations
- 60–75% of design team time goes to mechanical variations — not creative work
- Brand identity lives in a PDF that no system can enforce — the enforcement mechanism is a human reviewing every asset, every time
The challenge intensifies across physical substrates — embroidered fabric, backlit acrylic, neon tubing, 3D signage — where each material deforms brand geometry in fundamentally different ways that most AI systems ignore entirely.
graph TD
subgraph "Frontend SPA"
FE["LoRA Studio UI<br/>(Creator Zone + DS Zone)"]
end
subgraph "API Gateway"
GW["Gateway<br/>(Auth + Routing)"]
end
subgraph "Brand Management"
SVC_B["brand-service"]
SVC_P["persona-service"]
end
subgraph "Dataset Engineering"
SVC_I["ingest-service"]
SVC_C["caption-service"]
SVC_PK["package-service"]
end
subgraph "Training Orchestration"
SVC_CF["config-service (SSoT)"]
SVC_T["training-service"]
end
subgraph "Quality Assurance"
SVC_E["evaluation-service"]
SVC_R["review-service"]
end
subgraph "Deployment"
SVC_D["deploy-service"]
end
FE --> GW
GW --> SVC_B & SVC_P & SVC_I & SVC_C & SVC_CF & SVC_T & SVC_E & SVC_R & SVC_D
SVC_I --> SVC_C --> SVC_PK --> SVC_T --> SVC_E --> SVC_R --> SVC_D
SVC_B & SVC_P --> SVC_CF
SVC_R -->|Feedback Loop| SVC_C
15 microservices organized into 5 bounded contexts (DDD), communicating exclusively via REST API. Each service owns its database — zero shared state. The pipeline includes 3 Human-in-the-Loop checkpoints where automated processing pauses for expert review.
→ Deep dive: Full Architecture
The loss function in Flow Matching models:
has a profound architectural implication: when a visual attribute is present in the training image but absent from the caption, the model encodes it directly into the LoRA weights — permanently bound to the trigger word. This is not a heuristic; it is a deterministic effect of the mathematics.
- LOCKED attributes (logo geometry, typography): Never captioned → encoded in weights → permanent, unconditional
- UNLOCKED attributes (background, lighting): Always captioned → text-conditioned → flexible at inference
→ Deep dive: Mathematical Methodology
| Metric | Before | After | Improvement |
|---|---|---|---|
| Production Time | 6–8 weeks per campaign | Days | 85% reduction |
| Production Cost | $15K–40K per campaign | Fraction | 70% reduction |
| Brand Compliance | Manual review (human error) | Mathematical enforcement | 100% structural |
→ Deep dive: Evaluation & Business ROI
The system runs on NVIDIA GPUs (A100/H100/RTX PRO 6000) accessed via AWS EC2 instances. ComfyUI and Flux models are used in alignment with the NVIDIA Inception program. The architecture is cloud-agnostic — AWS is documented as the canonical reference with GCP and Azure equivalents mapped for each service.
| Layer | AWS (Canonical) | GCP (Equivalent) | Azure (Equivalent) |
|---|---|---|---|
| GPU Compute | EC2 G7E / P5 | Compute Engine A3 | NC A100 v4 / ND H100 v5 |
| Container Orchestration | ECS Express Fargate | Cloud Run | Container Apps |
| ML Platform | SageMaker | Vertex AI | Azure ML |
| Object Storage | S3 | Cloud Storage | Blob Storage |
→ Deep dive: Architecture (Cloud Layer)
This case study is based on a real production system. Evidence of implementation:
- MLV_Nodes_V3 — Open-source ComfyUI custom nodes (V3 API) used in the production pipeline: LoRA Stack, Ollama LLM, Dict Lookup, JSON Batcher
- OKLCH-Spectrum-Audit — Advanced OKLCH color palette visualizer demonstrating color science expertise
- Alpha-Compose — Professional image composition tool with 4K exports
- 10 public repositories + 37 private repositories covering the full ML pipeline stack
| Recognition | Source | Year |
|---|---|---|
| 🏆 Top 10 Prêmio Sebrae Startups (Winner: Media, Marketing & Advertising) | Agência Sebrae | 2025 |
| ☁️ AWS Case Study (via Select Soluções — AWS Partner) | Select Soluções | 2025 |
| 🧠 GenAI Producer (ACE Ventures + AWS classification) | Future Dojo | 2025 |
| 🏅 GenAI Awards 2024 — Transformative Use Cases in AI | Industry award | 2024 |
| 🌐 Web Summit Lisboa 2023 + Rio 2024 | International participation | 2023–24 |
| Section | What You'll Find | Time |
|---|---|---|
| paper/01-PROBLEM.md | The brand consistency gap, substrate challenge, market data | 5 min |
| paper/02-METHODOLOGY.md | Loss function math, LOCKED/UNLOCKED protocol, captioning pipeline | 10 min |
| paper/03-ARCHITECTURE.md | 15 microservices, DDD, REST contracts, cloud infrastructure | 15 min |
| paper/04-EVALUATION.md | 4 automated metrics, Human-in-the-Loop, business ROI | 5 min |
| paper/05-RELATED-WORK.md | Stability AI, Replicate, NVIDIA ecosystem, Brazilian GenAI landscape | 8 min |
| services/ | OpenAPI specs, SQL schemas, READMEs for 5 core services | 20 min |
| examples/ | UX flows for Creator onboarding, DS training monitor, review gallery | 10 min |
| GLOSSARY.md | 28+ ML terms explained for non-ML developers | 3 min |
Vandré Sales — AI/ML Architect specializing in multi-agent systems and generative AI for brand production. AWS CTO Fellowship alumni. Computer Engineering (UFG) + Physics (Unicamp).
This work is licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0). This is an architecture case study (documentation and diagrams), not runnable software. You are free to share and adapt with attribution.