AI Solutions Architect · Designing Idempotent, Secure & Observable Systems · Multi-Provider Orchestration & FinOps
I design AI-native systems that balance quality, compliance, and cost — from prototype to production-grade observability.
My focus is on RAG pipelines, multi-provider orchestration, and GDPR-compliant architectures across GCP, Azure, and AWS, built for idempotence, zero-trust, and measurable performance.
| Architecture Domain | Technologies & Methods |
|---|---|
| AI/ML Solutions | Multi-LLM architectures, prompt orchestration, RAG pipelines (LlamaIndex), semantic retrieval, embedding optimization |
| LLMOps & Observability | Langfuse / Helicone for tracing & metrics, Prometheus + Grafana dashboards, latency & cost telemetry, reproducible evals (RAGAS) |
| Data Engineering | PostgreSQL + pgvector, Qdrant, synthetic & masked datasets, lineage tracking, schema-aware ingestion |
| Cloud & Infrastructure | GCP / AWS / Azure, containerized runtimes (Docker / Compose), IaC with Terraform + Ansible, idempotent provisioning |
| Security & Compliance | Zero-Trust perimeter (mTLS, Cloudflare Tunnel), IAM least-privilege design, GDPR-by-design (Art. 25/30/32), auditable retention |
| FinOps & Optimization | Multi-provider orchestration (LiteLLM, ModelFusion, OpenRouter), rate-limiting, caching & graceful degradation, cost-aware routing |
| Agent Orchestration | LangGraph / CrewAI patterns, adaptive task routing, memory persistence, autonomous multi-agent coordination |
| Technical Documentation | ADRs, runbooks, solution architecture blueprints, compliance & funding documentation (EU / PT programs) |
🚀 NexoCLI
Lightweight CLI to automate developer workflows with pragmatic AI. Designed for reproducibility, short feedback loops, and minimal friction.
🤖 Nexo-Agents
Collection of 44 production-ready AI agent commands for Google AI CLI integration. Role-specific expertise with automatic context discovery.
Enterprise RAG Knowledge Platform
- Designed and deployed multi-stage retrieval (BM25 + dense + rerank) with >95% citation accuracy, ensuring traceable and explainable results.
- Implemented a privacy-by-design ingestion pipeline (GDPR Art. 25) with automated PII masking and audit-ready logging.
- PostgreSQL + pgvector self-hosted architecture, achieving €7.9 k cost reduction over 24 months compared to managed DB services.
- Integrated RAGAS-based evaluation harnesses for continuous model quality assurance and regression testing.
Multi-Provider LLM Gateway & Cost Optimization
- Architected LiteLLM-driven orchestration layer with 99.9 % uptime and automatic failover across Anthropic, OpenRouter, and Gemini providers.
- Applied caching and adaptive rate-limiting to achieve average cost ≈ $0.02 / 1 k tokens while maintaining consistent latency.
- Delivered end-to-end observability using Langfuse traces, Prometheus metrics, and Grafana dashboards linked to defined SLOs (P50 < 3 s | P95 < 8 s).
- Deployed under GCP infrastructure aligned with DNSH and renewable-energy compliance (>60 %), ensuring sustainability within EU funding standards.
- Consolidating RAG and LLMOps pipelines with multi-stage retrieval and rerankers, supported by RAGAS harnesses and Langfuse-based observability for trace-level evaluation.
- Advancing multi-provider orchestration through LiteLLM and LangGraph, introducing adaptive caching, load balancing, and real-time cost/latency optimisation.
- Extending serving efficiency via vLLM and TGI backends, exploring INT4 quantization and batch inference for scalable, low-latency deployment.
- Formalising IaC and FinOps blueprints to ensure reproducibility, auditability, and governance across hybrid and multi-cloud environments.
- Evolving compliance engineering with automated DPIA tooling, dynamic data masking, and continuous GDPR/AI Act alignment.
- Designing agent collaboration frameworks that enable self-diagnosing infrastructure and context-aware orchestration across pipelines.
| Domain / Layer | Implementation & Evidence |
|---|---|
| RAG Architecture | Multi-stage retrieval (BM25 + dense + rerank), LlamaIndex, hybrid scoring & evaluation with RAGAS, citation precision > 95%, context window optimization |
| LLM Orchestration | LiteLLM (multi-provider routing), ModelFusion / OpenRouter interoperability, 99.9 % uptime, cost ≈ $0.02 / 1 k tokens, adaptive load balancing |
| Agent Frameworks | LangGraph and CrewAI patterns, memory persistence, collaborative task routing, deterministic state machines |
| Compliance Engineering | GDPR Art. 25/30/32, DPIA templates, privacy-by-design ingestion, PII masking pipelines, encrypted audit logs |
| Observability & Telemetry | Langfuse / Helicone for inference tracing, Prometheus + Grafana dashboards, SLO enforcement (P50/P95 latency, uptime > 99.5 %), A/B eval tracking |
| Cloud & Infrastructure | GCP / AWS / Azure, Docker CE + Compose, Portainer CE, Cockpit, Terraform + Ansible, PostgreSQL + pgvector, optional Neo4j knowledge graphs |
| Automation & IaC | Declarative provisioning (Terraform), convergence & hardening (Ansible), idempotent rebuilds, policy-as-code validation |
| FinOps & Optimization | Multi-provider cost routing, rate limiting, caching, throughput forecasting, usage dashboards, per-tenant cost analytics |
| Development & Integration | Python 3.x, REST / WebSocket APIs, CLI tools, Git + GitHub Actions CI/CD, containerized development environments |
| Backup & DR | pg_dump automation, Duplicati incremental backups, GCP snapshots, RTO < 15 min validation scripts |
| Data Layer & Storage | PostgreSQL 16 + pgvector, MongoDB, Qdrant, synthetic & masked datasets, lineage & versioning |
| Documentation & Governance | Architecture Decision Records (ADRs), runbooks, Mermaid diagrams, Obsidian knowledge base, EU funding deliverables alignment |
- Website/Portfolio → https://en.nunosalvacao.pro/
- LinkedIn → https://www.linkedin.com/in/nsalvacao/
- Email → [email protected]
CET (EQF Level 5, PT). Unavailable during internship Sep–Dec 2025 (400h); open to opportunities thereafter.