AI/LLM Engineer e Platform Engineer: RAG, agentes e LLMOps — model serving com vLLM, Triton e NIM sobre Kubernetes multi-cloud (AWS · GCP · Azure), com observabilidade, FinOps e governança no mesmo pacote. Infraestrutura desde 1999: comecei com ASP, PHP e VBScript num servidor NT zumbindo do lado da mesa, anos antes da Locaweb existir.
Toque nos nodes brilhantes do grafo (edgeProxy ·
infra-operator · runner.codes) — são meus projetos
autorais. Os outros 18 nodes contam a stack que sustenta cada um deles.
Engenharia que vira código aberto. Os 4 nodes brilhantes do grafo são meus projetos autorais — clique neles direto pra inspecionar dentro da topologia, ou abra cada um aqui:
Tocador que mixa como DJ e anda sozinho — e o áudio toca no
navegador, não no servidor: o Go hexagonal só mantém a
agenda e decide a próxima faixa. Arquivos locais, Spotify e YouTube na
mesma fila, com crossfade real entre faixas, BPM medido por
ffmpeg/aubio e versão SEM-VOZ separada por IA
(Demucs) — na CPU de graça ou numa GPU alugada por ~US$ 0,001 a faixa.
Código aberto sob MIT em github.com/andrebassi/autodj, rodando em autodj.andrebassi.com.br.
Proxy TCP/HTTP distribuído escrito em Rust para edge computing.
Routing geo-aware via GeoIP + Anycast, terminação TLS por região,
observabilidade OTel nativa, deploy multi-PoP (GRU · IAD · SIN · FRA).
Substitui o combo nginx + haproxy + lua por um único
binário com ~3 MB de RSS e p99 sub-milissegundo no L4.
Operator Kubernetes em Go que declara recursos AWS (EKS, RDS, S3, IAM, VPC, Route53) como CRDs nativos. Reconciliação contínua, drift detection, rollback automático, zero clique no console. Move parte do Terraform pra dentro do loop de controle do K8s — IaC vira Runtime-AC.
Runtime seguro e isolado para código gerado por LLM. microVMs Firecracker (KVM puro) em vez de containers — ~125 ms cold-boot, snapshot/restore sub-segundo, 40+ runtimes prontos (Python, Node, Go, Rust, Bash, Java, .NET, Ruby…). MCP-ready: Claude/GPT/Cursor escrevem, runner executa, humano valida.
A mesma linha de 1999 a 2026, vista de três ângulos: as eras de tecnologia que atravessei, os cargos que ocupei e as camadas de stack que mantenho. A régua abaixo é compartilhada — aponte qualquer item e as outras duas leituras acendem no mesmo intervalo.
Cada linha é uma camada que sobreviveu ao tempo — adicionada quando virou produção, mantida porque ainda paga conta. No topo, a camada de IA; embaixo, a plataforma que segura ela de pé.
RAG com busca híbrida e re-ranking · grounding em fontes autorizadas · guardrails e mitigação de alucinação. Agentes com function calling, contratos de saída em JSON Schema e validação estrutural · MCP · embeddings/vector DB · Vertex AI · Bedrock · OpenRouter · Claude · GPT · Cursor.
Fine-tuning de modelos abertos com LoRA e QLoRA (Unsloth · Axolotl) nas famílias Llama · Qwen · Phi · Kimi. Model serving com vLLM · NVIDIA Triton · NIM · ONNX. MLflow · Kubeflow · Ray. GPU no K8s (device plugin · MIG) · Jetson Orin NX · runtimes Firecracker pra código LLM-gerado.
Go · Rust · Python · TypeScript · Bash. Antes: C# · .NET · VB · Ruby · Java/Spring · PHP · ASP.
Docker · EKS · GKE · AKS · OCI · k3s · k0s · MicroK8s · Talos · OrbStack · Helm · Kaniko.
Terraform · Terragrunt · Ansible · Pulumi · módulos reusáveis · multi-cloud · bare metal · Proxmox.
Istio · Cilium · Kong · Nginx · HAProxy · Keycloak · WireGuard · OpenVPN · BGP · Cloudflare.
GitLab · GitHub Actions · Jenkins · Tekton · ArgoCD · FluxCD · Renovate · SAST · DAST · IAST · Trivy.
Prometheus · Grafana (LGTM) · OTel · ClickHouse · Datadog · Dynatrace · NewRelic · KEDA-driven SLOs.
Rightsizing · spot/preemptible · budget por workload de GPU. RBAC · IAM · OIDC/mTLS · secrets management (External Secrets · Vault) · LGPD · auditoria · ADR. ITSM com GLPI publicado por Cloudflare Zero Trust e SSO Microsoft 365.
Linux (kernel · cgroup · KVM) · PostgreSQL · Redis · ClickHouse · MongoDB · KubeBlocks. Mensageria com Kafka e RabbitMQ para arquiteturas orientadas a eventos.
Oito cargos, 27 anos, do plantão de servidor NT à plataforma de IA. Aqui é o que eu entreguei, e para quem — a era de tecnologia de cada período está na lente eras.
Da ASP num NT4 zumbindo do lado da mesa a microVMs Firecracker rodando código LLM-gerado. Cada era adicionou uma camada — nada foi descartado por moda.
ASP, PHP e VBScript eram os bloquinhos pra web. Servidor NT e Linux zumbindo do lado da mesa — anos antes da Locaweb existir.
.NET 1.0 (fev/2002) virou o kit de apps web e desktop. Em paralelo, AWS (2006), GCP (2008) e Azure (2010) abriram a porteira da nuvem.
Ruby, JSP e Spring Boot eram as ferramentas da época. Apache servindo HTTP, Tomcat hospedando servlets — eu administrava os dois.
Docker empacotou cada app em container. Kubernetes (v1.0, jul/2015) organizou e escalou tudo na nuvem. Em produção desde 2017 — EKS · GKE · OCI · Talos.
Transformer (2017) e BERT (2018) entenderam linguagem. Vision (CNN → ViT) decifrou imagens. RAG (mai/2020) aprendeu a buscar contexto antes de responder.
Copilot e ChatGPT (2022) abriram o caminho. Agora, agents geram código, redigem PRDs e tocam o fluxo inteiro.
Consultoria em AI Platform: RAG e agentes em produção, model serving (vLLM · Triton · NIM) em Kubernetes multi-cloud, com observabilidade, FinOps e governança. edgeProxy serve, infra-operator orquestra, runner.codes executa o código gerado — humano valida.
Consultoria em AI Platform · LLMOps & model serving · plataformas Kubernetes · mentoria · open source.
© 2027 · OSI model · canvas 2D · gsap · zero build step
AI/LLM Engineer and Platform Engineer: RAG, agents and LLMOps — model serving with vLLM, Triton and NIM on multi-cloud Kubernetes (AWS · GCP · Azure), with observability, FinOps and governance in the same package. Infrastructure since 1999: I started with ASP, PHP and VBScript on an NT server humming next to my desk, years before Locaweb existed.
Tap the glowing nodes in the graph (edgeProxy ·
infra-operator · runner.codes) — those are my own
projects. The other 18 nodes tell the stack that supports each one.
Engineering that becomes open source. The 4 glowing nodes in the graph are my own projects — click them right inside the topology, or open each one here:
A player that mixes like a DJ and runs unattended — and the audio plays
in the browser, not on the server: the hexagonal Go
backend only keeps the schedule and decides what comes next. Local
files, Spotify and YouTube in one queue, with real crossfade, BPM
measured by ffmpeg/aubio and an AI-separated
vocal-free version (Demucs) — free on CPU, or ~US$ 0.001 per track on a
rented GPU. Open source under MIT at
github.com/andrebassi/autodj, live at
autodj.andrebassi.com.br.
Distributed TCP/HTTP proxy written in Rust for edge computing.
Geo-aware routing via GeoIP + Anycast, per-region TLS termination,
native OTel observability, multi-PoP deploy (GRU · IAD · SIN · FRA).
Replaces the nginx + haproxy + lua combo with a single
binary at ~3 MB RSS and sub-millisecond p99 at L4.
Kubernetes operator in Go that declares AWS resources (EKS, RDS, S3, IAM, VPC, Route53) as native CRDs. Continuous reconciliation, drift detection, automatic rollback, zero console clicks. Moves part of Terraform into the K8s control loop — IaC becomes Runtime-AC.
Secure, isolated runtime for LLM-generated code. Firecracker microVMs (pure KVM) instead of containers — ~125 ms cold-boot, sub-second snapshot/restore, 40+ ready runtimes (Python, Node, Go, Rust, Bash, Java, .NET, Ruby…). MCP-ready: Claude/GPT/Cursor write, runner executes, a human validates.
The same 1999 → 2026 line seen from three angles: the technology eras I went through, the roles I held and the stack layers I keep running. The ruler below is shared — hover any item and the other two readings light up over the same span.
Each line is a layer that survived time — added when it hit production, kept because it still pays the bills. AI on top; underneath, the platform that holds it up.
RAG with hybrid search and re-ranking · grounding on authorized sources · guardrails and hallucination mitigation. Agents with tool orchestration via function calling, output contracts in JSON Schema and structural validation · MCP · embeddings/vector DB · Vertex AI · Bedrock · OpenRouter · Claude · GPT · Cursor.
Fine-tuning of open models with LoRA and QLoRA (Unsloth · Axolotl) across the Llama · Qwen · Phi · Kimi families. Model serving with vLLM · NVIDIA Triton · NIM · ONNX. MLflow · Kubeflow · Ray. GPU on K8s (device plugin · MIG) · Jetson Orin NX · Firecracker runtimes for LLM-generated code.
Go · Rust · Python · TypeScript · Bash. Before: C# · .NET · VB · Ruby · Java/Spring · PHP · ASP.
Docker · EKS · GKE · AKS · OCI · k3s · k0s · MicroK8s · Talos · OrbStack · Helm · Kaniko.
Terraform · Terragrunt · Ansible · Pulumi · reusable modules · multi-cloud · bare metal · Proxmox.
Istio · Cilium · Kong · Nginx · HAProxy · Keycloak · WireGuard · OpenVPN · BGP · Cloudflare.
GitLab · GitHub Actions · Jenkins · Tekton · ArgoCD · FluxCD · Renovate · SAST · DAST · IAST · Trivy.
Prometheus · Grafana (LGTM) · OTel · ClickHouse · Datadog · Dynatrace · NewRelic · KEDA-driven SLOs.
Rightsizing · spot/preemptible · per-workload GPU budget. RBAC · IAM · OIDC/mTLS · secrets management (External Secrets · Vault) · LGPD · audit trails · ADRs. ITSM with GLPI published through Cloudflare Zero Trust and Microsoft 365 SSO.
Linux (kernel · cgroup · KVM) · PostgreSQL · Redis · ClickHouse · MongoDB · KubeBlocks. Messaging with Kafka and RabbitMQ for event-driven architectures.
Eight roles, 27 years, from NT server on-call to AI platform. This is what I delivered, and for whom — the technology era behind each period lives in the eras lens.
From ASP on an NT4 humming next to the desk to Firecracker microVMs running LLM-generated code. Each era added a layer — nothing was discarded for fashion.
ASP, PHP and VBScript were the building blocks of the web. NT and Linux servers humming next to the desk — years before Locaweb existed.
.NET 1.0 (Feb/2002) became the kit for web and desktop apps. Meanwhile, AWS (2006), GCP (2008) and Azure (2010) opened the cloud gates.
Ruby, JSP and Spring Boot were the tools of the era. Apache serving HTTP, Tomcat hosting servlets — I ran both.
Docker packed every app into a container. Kubernetes (v1.0, Jul/2015) organized and scaled it all in the cloud. In production since 2017 — EKS · GKE · OCI · Talos.
Transformer (2017) and BERT (2018) understood language. Vision (CNN → ViT) decoded images. RAG (May/2020) learned to fetch context before answering.
Copilot and ChatGPT (2022) paved the way. Now agents generate code, write PRDs and drive the whole flow.
AI Platform consulting: RAG and agents in production, model serving (vLLM · Triton · NIM) on multi-cloud Kubernetes, with observability, FinOps and governance. edgeProxy serves, infra-operator orchestrates, runner.codes executes the generated code — a human validates.
AI Platform consulting · LLMOps & model serving · Kubernetes platforms · mentoring · open source.
© 2027 · OSI model · canvas 2D · gsap · zero build step