AI Engineer Β· Privacy-First RAG & Multi-Agent Systems Β· GDPR-Compliant Generative AI
Building enterprise AI infrastructure that runs locally, scales globally, and leaks nothing.
I engineer Generative AI platforms for highly-regulated environments, ensuring data privacy, scalable orchestration, and independence from cloud LLM providers. The market demands AI that respects GDPR β I build systems where infrastructure sovereignty is the default.
Instead of just prototyping, I approach AI Engineering with a results-driven mindset:
- Sovereign Enterprise AI: Companies need AI without risking proprietary data leaks. Result: Architected systems like Codex One and IA News Agent, utilizing open-weights models (Ollama) and on-prem ChromaDB to achieve 100% data sovereignty and zero external API dependencies, making them audit-ready for strict compliance.
- Workflow Automation at Scale: Manual data processing bottlenecks business growth. Result: Built JobMatch, a multi-agent RAG platform that parses unstructured resumes and performs semantic ranking, reducing human evaluation time by 95% while retaining precise skill-matching.
- High-Throughput & Low-Latency AI: Cloud API costs scale linearly; local inference requires optimization. Result: Implemented batch prompting, asynchronous LangGraph orchestration, and exact/semantic dual-layer caching (achieving 12ms inference latency in Visual Tagger) to push CPU/GPU hardware to its limits.
Privacy-First RAG Pipelines Β· Multi-Agent Orchestration Β· Data Engineering (Spark + Scala)
Local LLM Deployment Β· Multi-Cloud Infrastructure Β· Security & GDPR Compliance
π° IA News AgentAutonomous multilingual intelligence pipeline Aggregates, enriches, and validates AI research across 10 languages and 7 source types. Uses a dual-model LLM strategy (Qwen 3B + 14B) driven by a stateful LangGraph reasoning core. Complete on-prem execution with zero data leakage.
|
π Codex OnePrivacy-first enterprise knowledge base β fully offline RAG Zero-egress architecture with pre-ingestion Threat Quarantine and automated PII Scrubbing. Uses CPU-bound ONNX embeddings and multilingual Cross-Encoder reranking (FlashRank) to isolate GPUs exclusively for Local LLMs (Ollama), delivering audit-ready compliance and high-performance retrieval.
|
π€ JobMatchMulti-agent RAG platform for HR automation Automated talent screening reducing manual evaluation by 95%. Semantic search over unstructured data, dynamic ranking, LLM-agnostic inference (OpenAI / Gemini / local). Real-time analytics via WebSockets.
|
π·οΈ Visual TaggerHigh-performance AI image classification β 12ms latency Concurrent ViT + CLIP Zero-Shot inference with confidence-score aggregation. Parallel model execution for automated multi-label tagging at production scale.
|
π Cosmetic CosmosAI-augmented scalable sales management platform Distributed high-performance ScyllaDB backend handling concurrent retail transactions. Features an embedded role-aware local LLM assistant (Ollama) providing contextual KPIs and mentorship without exposing metrics to third parties.
|
π§± Systems ProgrammingLow-level engineering β C from scratch Custom standard library, Unix process pipelines, 2D rendering engine. Built at 42 SΓ£o Paulo with zero external dependencies β raw memory management, file descriptors, and bitwise ops.
|
flowchart TD
%% Modern Premium Styling
classDef gateway fill:#1E293B,stroke:#475569,stroke-width:2px,color:#F8FAFC,rx:8,ry:8
classDef orchestrator fill:#334155,stroke:#64748B,stroke-width:2px,color:#F8FAFC,rx:8,ry:8
classDef security fill:#991B1B,stroke:#EF4444,stroke-width:2px,color:#FEF2F2,rx:8,ry:8
classDef pipeline fill:#0F766E,stroke:#14B8A6,stroke-width:2px,color:#F0FDFA,rx:8,ry:8
classDef database fill:#4338CA,stroke:#6366F1,stroke-width:2px,color:#EEF2FF,rx:8,ry:8
classDef inference fill:#6B21A8,stroke:#A855F7,stroke-width:2px,color:#FAF5FF,rx:8,ry:8
subgraph Edge ["π API & Gateway Layer"]
direction LR
Client(["Client Apps"])
API["API Gateway<br/><i>(Auth & Rate Limits)</i>"]
end
subgraph Security ["π‘οΈ Enterprise Security"]
direction TB
Threat["Threat Scanner<br/><i>(Prompt Injection / Quarantine)</i>"]
PII["Compliance Agent<br/><i>(PII Scrubber & Masking)</i>"]
end
subgraph Core ["π Two-Stage RAG Pipeline"]
direction TB
Router{"Semantic Router"}
Embed["CPU Embeddings<br/><i>(FastEmbed / ONNX)</i>"]
Retrieve["Initial Retrieval<br/><i>(Top-K Similarity)</i>"]
Rerank["Cross-Encoder Reranker<br/><i>(FlashRank / High Precision)</i>"]
end
subgraph Storage ["πΎ Data Sovereignty Layer"]
direction LR
Redis[("Redis<br/><i>(Session Memory)</i>")]
Vector[("ChromaDB<br/><i>(Local Vector Store)</i>")]
end
subgraph Engine ["π§ Local Inference Engine"]
direction LR
LLM("Open-Weights LLMs<br/><i>(Ollama / Zero Data Egress)</i>")
end
%% Flow Connections
Client ===>|Requests| API
API --->|Dispatch| Router
%% Ingestion Flow
Router -.->|Document Upload| Threat
Threat -.->|Safe Data| PII
PII -.->|Sanitized Text| Embed
Embed -.->|Upsert Vectors| Vector
%% Query Flow
Router ===>|Query| Embed
Embed ===>|Dense Vectors| Retrieve
Retrieve <===>|Fetch Top 20| Vector
Retrieve ===>|Pass Chunks| Rerank
Rerank ===>|Top 5 Refined| LLM
%% State
Router <-->|Maintains Context| Redis
%% Styling Application
class Edge,Client,API gateway
class Security,Threat,PII security
class Core,Router,Embed,Retrieve,Rerank pipeline
class Storage,Redis,Vector database
class Engine,LLM inference
|
AI & ML |
Data Engineering |
Cloud & DevOps |
Languages & APIs |
- π Privacy & Law by Design: Engineered for environments where strict European GDPR dictates are mandatory. Implementing Offline LLMs and on-prem vector databases to guarantee zero third-party exposure.
- π Vendor-Agnostic Architecture: Building decoupled, LLM-agnostic platforms. Codebases designed such that swapping OpenAI β Gemini β Local Ollama requires zero restructuring.
- π Scalability & Resiliency: Designing hybrid architecturesβfrom 200GB/day distributed Apache Spark systems to ultra-lightweight Dockerized microservices wrapped in circuit breakers and semantic cache layers.
- βοΈ First-Principles Thinking: Grounded in the brutal 42 SΓ£o Paulo methodology (C, Unix, memory management). I don't just glue APIs together; I profoundly understand the performance parameters executing them.
42 SΓ£o Paulo β Software Engineering (Peer-to-peer, zero-professor model)
Unisinos β Analysis & Systems Development
Bayswater College, London β Exchange Program
π₯ Moving The Cities β Unisinos Γ SAP Γ FH MΓΌnster (Germany) Γ UAS7
π GeraΓ§Γ£o Caldeira β First Class, Institute Caldeira