AI/ML Architect specializing in Multi-Agent Systems, Generative AI, and Production ML Infrastructure.
Enterprise clients like Porto Seguro (15M customers) and RD Saúde/Drogasil needed to scale personalized visual asset production without 5× their creative teams. I architected a platform that orchestrates the entire lifecycle — from brand ingestion to mass generation via API — with clear separation of concerns: security in data handling, efficiency in GPU processing, and governance across the pipeline.
flowchart TB
subgraph INGEST["📥 Brand Ingestion"]
A1["Brandbook Upload<br/>(colors, fonts, guidelines)"] --> A2["Visual Identity Analysis<br/>(logo topology, palette extraction)"]
A2 --> A3["Persona Definition<br/>(demographics, attributes, rules)"]
end
subgraph DATASET["🗂️ Proprietary Dataset Engine"]
B1["Automated Image Curation<br/>(face crop, resize, quality filter)"] --> B2["AI Captioning Pipeline<br/>(Vision LLM + LOCKED/UNLOCKED protocol)"]
B2 --> B3["Validation & Packaging<br/>(violation scan, 1024px Lanczos, ZIP)"]
end
subgraph TRAINING["🧬 LoRA Fine-Tuning Factory"]
C1["Identity LoRAs<br/>(persona faces — Flux.1 DiT)"] --> C2["Style LoRAs<br/>(brand architecture patterns)"]
C2 --> C3["Color Palette LoRAs<br/>(chromatic consistency)"]
C3 --> C4["Loss Monitoring & EMA<br/>(bf16, flowmatch, adaptive rank)"]
end
subgraph INFERENCE["⚡ Generation Engine"]
D1["ComfyUI Workflow DAG<br/>(Custom Nodes V3 API)"] --> D2["Multi-LoRA Orchestration<br/>(identity + style + color stacking)"]
D2 --> D3["LLM Prompt Enhancement<br/>(scene description enrichment)"]
D3 --> D4["GPU Rendering<br/>(H100 / A100 / Blackwell)"]
end
subgraph DELIVERY["🚀 Enterprise Delivery"]
E1["REST API Gateway<br/>(FastAPI + BentoML containers)"] --> E2["Multi-Cloud Deploy<br/>(AWS ECS + GCP Cloud Run)"]
E2 --> E3["SaaS Interface<br/>(React + WASP framework)"]
E3 --> E4["Batch Generation<br/>(high-volume, brand-compliant)"]
end
INGEST --> DATASET --> TRAINING --> INFERENCE --> DELIVERY
mindmap
root((Vandré Sales<br/>AI/ML Architect))
🧠 Generative AI
LLM Fine-Tuning
LoRA / QLoRA
ai-toolkit
kohya_ss
Replicate API
RAG Pipelines
LangChain
LlamaIndex
Vector DBs
Multi-Agent Systems
Orchestration
Tool Calling
Memory Management
Computer Vision
Flux.1 / Flux.2
SDXL
ComfyUI V3 Nodes
☁️ Cloud & Infrastructure
AWS
SageMaker
Bedrock
ECS Express
EC2 GPU
GCP
Vertex AI
AI Studio
Cloud Run
GPU Operations
H100 / A100
Blackwell RTX PRO
CUDA / cuDNN
💻 Full-Stack Development
Backend
Python / FastAPI
BentoML
Docker / K8s
Frontend
TypeScript / React
Next.js / Vite
Tailwind / Shadcn
SaaS Platform
WASP Framework
Prisma ORM
Stripe Integration
📐 Architecture & Governance
Spec-Driven Development
Cognitive Shell
AI Governance Framework
Semantic Versioning
Design Systems
Atomic Design
Shadcn Variants
Component Libraries
When building LoRAs for enterprise persona identity at scale, I discovered the bottleneck isn't training — it's dataset preparation. A poorly written caption destroys LoRA consistency. I developed a proprietary end-to-end process with a conditional captioning protocol (LOCKED/UNLOCKED) that mathematically guarantees permanent attributes are absorbed by the trigger word while variable attributes remain prompt-controllable.
sequenceDiagram
participant SRC as 📸 Source Images
participant PREP as 🔧 Curation
participant CAP as 🏷️ AI Captioning
participant LOCK as 🔒 Conditional Protocol
participant QA as ✅ Quality Gate
participant TRAIN as 🧬 Training
participant EVAL as 📊 Evaluation
participant PROD as ⚡ Production
SRC->>PREP: Raw images collected
PREP->>PREP: Crop, resize, diversity audit
PREP->>CAP: Clean image set
CAP->>CAP: Vision LLM auto-captioning
CAP->>LOCK: Raw captions
LOCK->>LOCK: Classify attributes
Note over LOCK: LOCKED → never describe<br/>→ learned by trigger word
Note over LOCK: UNLOCKED → always describe<br/>→ controllable by prompt
LOCK->>QA: Conditioned captions
QA->>QA: Violation scan + consistency audit
QA->>TRAIN: Validated dataset
TRAIN->>TRAIN: LoRA fine-tuning (Flux.1 DiT)
TRAIN->>EVAL: Model checkpoints
EVAL->>EVAL: Trigger activation test
EVAL->>EVAL: Attribute controllability test
EVAL-->>TRAIN: Feedback loop if needed
EVAL->>PROD: Production-ready LoRA
PROD->>PROD: Multi-LoRA deployment
PROD->>PROD: Enterprise API serving
Coordinating multiple AI projects simultaneously (SaaS, GPU infra, LoRA training, APIs), I realized the biggest risk wasn't technical — it was cognitive entropy between work sessions with AI agents. Each session started from zero: no memory, no context, no governance. I created a spec-driven agentic development framework where every action is preceded by formal specification, validated by adversarial QA, and versioned with full traceability. The AI agent operates as runtime; the framework governs.
flowchart TB
subgraph SPEC["📋 Specification Layer"]
S1["🔍 Dossier<br/>(forensic investigation)"] --> S2["💡 Concept<br/>(architectural ideation)"]
S2 --> S3["📐 Plan<br/>(tactical strategy)"]
S3 --> S4["📝 Steps<br/>(granular execution guide)"]
end
subgraph QA["🛡️ Quality Assurance"]
Q1["😈 Devil<br/>(adversarial stress test)"] --> Q2{Approved?}
Q2 -->|Yes| Q3["✅ APPROVED<br/>with restrictions"]
Q2 -->|No| Q4["🔄 Rework<br/>back to Plan"]
end
subgraph EXEC["⚡ Execution Layer"]
E1["🤖 AI Agent Runtime<br/>(MCP-connected tools)"] --> E2["🔧 Tool Orchestration<br/>(read, write, execute, search)"]
E2 --> E3["📊 Telemetry<br/>(state tracking documents)"]
end
subgraph GOV["🏛️ Governance Layer"]
G1["📜 Constitution<br/>(immutable laws)"]
G2["📦 Semantic CLI<br/>(cognitive command shell)"]
G3["🔄 Versioning<br/>(SemVer + Changelogs)"]
end
S4 --> Q1
Q3 --> E1
Q4 --> S3
GOV -.->|governs| SPEC
GOV -.->|governs| QA
GOV -.->|governs| EXEC
timeline
title Technology & Innovation Journey — Vandré Sales
1990 : 💻 Programming Instructor (age 14)
: COBOL 80, FORTRAN, Algorithms
: Youngest CS teacher in Brasília
1998 : 🔬 Physics & Quantum Mechanics
: Unicamp — relativistic physics
: 4000+ students across 8 institutions
2003 : 🏫 Educational Technology Director
: ESAMC — pedagogical systems
: Robotics lab + digital infrastructure
2010 : 🚀 Innovation & Design Thinking
: PontoGet — corporate innovation consultancy
: UC Berkeley Design Thinking certified
2013 : 📊 Startup Founder — Consumer BI
: Tippz — real-time analytics platform
: Pitched to Sequoia, IBM, NASA (SF 2015)
: Tel Aviv Stock Exchange pitch (2016)
2016 : 🏦 Startup Exit → Itaú
: Tippz acquired by Brazil's largest bank
: ACE Accelerator — Head of Hub Goiânia
2021 : 🤖 AI-First SaaS Platform
: Meliva.ai — GenAI content orchestration
: Multi-model, multi-cloud architecture
2024 : ☁️ AWS CTO Fellowship + Accelerators
: Dr. Werner Vogels' global program
: Google for Startups + Microsoft Founders + NVIDIA Inception
2025 : 🏆 Awards & Global Recognition
: NVIDIA Top 12 Startups (Re:Invent Las Vegas)
: Sebrae TOP 10 National (3167 companies)
: Web Summit speaker (Lisbon, Rio)
2026 : 🌍 Enterprise Scale + Global Speaker
: Beijing HICOOL Summit speaker
: Production GPU infra (H100, Blackwell)
: 41 repositories — full E2E AI pipeline
| Project | Description | Tech |
|---|---|---|
| LoRA Studio — Architecture Case Study ⭐ | AI-Powered Brand Consistency at Scale — 15 microservices, LOCKED/UNLOCKED protocol, Flow Matching math. AWS Case Study Partner · Top 12 Worldwide Startups 2025 NVIDIA / AWS Re:Invent | Architecture / GenAI |
| Blueprint-RAG-Agnostic ⭐ | A highly modular, LLM-agnostic & Vector DB-agnostic RAG architecture featuring a 1-Click Interactive TUI Wizard for zero-drift template deployments. | Python / LangChain |
| Alpha-Compose | Precision image orchestrator — compose subjects over backgrounds and batch-export up to 4K | TypeScript |
| OKLCH-Spectrum-Audit | Advanced OKLCH palette visualizer with HEX conversion, luminance audit, and CSS export | TypeScript |
| batch-image-crop | Batch image cropping tool with aspect ratio presets and ZIP export up to 4K | TypeScript |
| 41 repositories | Full pipeline: Dataset→LoRA→Inference→API→Product | Multi-lang |
- 🔗 LinkedIn — Open to opportunities
- 🌐 meliva.ai — AI-powered content platform
- 📧 vandre.sales@gmail.com
"Credentials without visibility are like code without deploy — they exist, but generate no value."