CAPSTONE, Staff/Principal Specialization Track

1. Problema de Engenharia

Senior generalista entrega Logística v3 distribuído. Staff/Principal escolhe um eixo de profundidade e produz output que demonstra mastery suficiente pra título Staff em empresa séria. Não é mais um capstone uniforme, é o ponto onde a trajetória se ramifica.

Você já fez:

• 05-01: 2 toys de baixo nível (build-from-scratch).
• 05-02: 3 capstones de domínios distintos.
• 05-03: proposta org + estudo de caso.
• 05-04: paper habit + 1 paper implementado.
• 05-05: stream público com 6+ posts + talk.
• 05-06: 3 mentees acompanhados.
• 05-07 opcional: tracker IoT.

Esse capstone pega tudo isso e cristaliza num showcase de specialization + promo case público. Output não é projeto novo; é integração + posicionamento do que você já fez.

Tracks recomendados:

• A. Distributed Systems Engineer: deep em consensus, replication, multi-region. Owner de toy DB (05-01) + Raft (05-04 implementado) + paper writing.
• B. Platform / Infra Engineer: K8s, observability, DevEx. Toy scheduler (05-01) + plataforma interna (Logística v2 platform team angle) + multi-cluster.
• C. Frontend / DX Architect: React deep, build tools, framework design. Toy compiler/runtime (05-01) + DX library (04-15).
• D. Data / ML Engineer: pipelines, feature stores, RAG/LLM. 05-02 RAG capstone + 04-13 streaming + model serving.
• E. Security Engineer: AppSec, crypto, threat modeling. 01-12 base + 03-08 + 04-11 web3 + secure design reviews.
• F. Founding / Product Engineer: range generalist + business sense. 04-16 unit economics + 05-03 org + 05-05 public output.
• G. AI Infrastructure Engineer: GPU clusters, inference servers, training pipelines, evals, vector DBs em prod. 04-10 + 05-02 + custom infra.

Pick 1. O capstone produz portfolio e promo case alinhados.

---

2. Teoria Hard

2.1 Specialization vs generalist

Staff pode ser super-deep specialist (Distributed Systems Engineer @ Stripe) ou super-generalist (Tech Lead spanning N domains). Ambos válidos.

Specialist: invest 70% no eixo, 30% breadth. Generalist: 50/50 ou even 30/70 depth/breadth.

Empresas grandes (Google, Meta, Stripe) têm slot pra ambos. Startups + scale-ups privilegiam generalists em IC.

Não-decisão default = generalist. Active choice = specialist.

2.2 Promo case: anatomia

Promo case (brag doc) é doc 5-10 páginas + supporting artifacts:

• Scope crescente cobrindo 12-18 meses.
• Impact em $$ / users / engineering velocity.
• Influence: docs, talks, mentees grown.
• External validation: blog reach, OSS stars, conf talks.
• Case studies de problemas resolvidos.

Sponsor (manager senior) defende em comitê.

2.3 External validation

Internal promo + external visibility = strong signal. Patterns:

• Blog post conhecido em area.
• Talk em conf reconhecida.
• OSS lib usado por empresas third-party.
• Mentee promoted publicly.
• Cited in industry write-up.

Não pra ego, pra validate scope além de empregador atual.

2.4 Job market posicionamento

Staff offer numbers em Brasil 2026: $150-300k+ TC (USD/year-equiv) tier 1; menores em mid-market. Comp depende de:

• Location (remote SF wage compression).
• Equity vs cash mix.
• Stage of company.
• Specialization scarcity.

Don't optimize comp short-term. Optimize career mobility.

2.5 Track A, Distributed Systems Engineer

Output:

• Toy DB completo + toy queue (05-01).
• Raft implementation (05-04 paper).
• TLA+ specs em produção (04-14).
• Paper read + 1 implemented + 1 blog post deep.
• Logística v3 production-grade distributed.
• Talk em meetup/conf sobre algum aspecto.
• Lib OSS (idempotency-kit ou similar).

Companies que valorizam: Stripe, Cockroach, Confluent, AWS, Google Cloud, Cloudflare, Anthropic.

2.6 Track B, Platform / Infra Engineer

Output:

• Toy scheduler / load balancer (05-01).
• Platform team proposal (05-03) implementada na Logística (or thought experiment).
• Multi-cluster K8s setup com IaC.
• Internal DevEx improvements measured (deploy time, failed deploy rate).
• Observability stack rolled out.
• Talk sobre platform building.
• Lib OSS pra DevEx.

Companies: Vercel, Railway, Fly.io, GitHub, GitLab, AWS, Heroku-style platforms.

2.7 Track C, Frontend / DX Architect

Output:

• Toy compiler/runtime (05-01 + 01-13).
• Framework lib (similar ao TanStack scope).
• Build tool optimization caso real.
• React/Next deep contribution OSS.
• DX-focused blog series.
• Talk em React/Next/Astro/Svelte conf.

Companies: Vercel, Netlify, Linear, Figma, Notion, Stripe.

2.8 Track D, Data / ML Engineer

Output:

• 05-02 RAG capstone production-quality.
• ST13 pipeline streaming + batch.
• Feature store / vector DB lib.
• Model serving infra (low-latency).
• Eval framework + dataset curated.
• Blog series sobre MLOps.
• Talk em MLOps/data conf.

Companies: Databricks, Pinecone, OpenAI, Anthropic, Hugging Face, Weights & Biases.

2.9 Track E, Security Engineer

Output:

• Threat models documentados (Logística + capstones).
• Pentest report Logística com remediations.
• 01-12 cripto deep dive blog.
• Web3 contract audit (04-11 capstone).
• Bug bounty findings (Hackerone profile).
• Talk em OWASP / DefCon-adjacent.

Companies: Cloudflare, Trail of Bits, Doyensec, Latacora, Anthropic.

2.10 Track F, Founding / Product Engineer

Output:

• Logística como side product testado com 5+ users reais.
• Unit economics simulator + dashboards (04-16).
• Org proposal com customer-discovery framing.
• Public output focused on bridging engineering + product.
• Talk em founder-engineering oriented event.
• Possibly small product launched (Plasmic-style).

2.11 Track G, AI Infrastructure Engineer

Categoria que explodiu em 2024-2026. Diferente de "ML engineer" tradicional (treinar modelos) ou "AI app dev" (consume API): infraestrutura que serve modelos em escala: GPU clusters, inference servers, training pipelines, vector DBs em prod, evals em CI.

Output:

• Inference cluster próprio (toy ou real): vLLM ou TGI servindo modelo open-weight (Llama 3, Mistral) com batch dynamic, KV cache. Mede tokens/sec, p99 latency, GPU utilization.
• Training pipeline pequeno: fine-tune LoRA num modelo base sobre dataset curado. Track via Weights & Biases ou MLflow. Documenta loss curves, eval metrics, decisions.
• Vector DB em produção: pgvector ou Qdrant servindo RAG real (continuação do 05-02). Embedding strategy documentada, recall metrics, cost/QPS analysis.
• Eval framework: golden dataset + LLM-as-judge + A/B em produção. Report comparando 3+ modelos em metric específico do seu domínio.
• GPU cost optimization case study: análise de spot vs on-demand, batch vs streaming, caching de KV, quantization (INT8/FP8). Antes vs depois com números.
• Observability LLM: tracing de prompts, token usage tracking, cost-per-customer, drift detection.
• Blog series MLOps: 4-6 posts sobre challenges reais. "Why we moved from OpenAI to self-hosted Llama" tipo de artigo.
• Talk em Ray Summit, MLOps Community, AI Engineer Summit.

Stack típico:

• Serving: vLLM, TGI (HuggingFace), TensorRT-LLM, Triton.
• Training: PyTorch, JAX, Lightning. Distributed: DeepSpeed, FSDP.
• Vector DB: pgvector, Qdrant, Pinecone, Weaviate, Milvus.
• Orchestration: Ray, Kubeflow, Argo Workflows.
• Monitoring: Langfuse, Helicone, Phoenix, Arize.
• Hardware: NVIDIA H100/H200, AMD MI300, Google TPU v5. Cloud: AWS p5/p6, GCP 02-03, Lambda Labs, RunPod, Modal.

Companies: OpenAI, Anthropic, Mistral, Together, Replicate, Modal, Hugging Face, Databricks, NVIDIA, Cohere, AI startups Series A-C.

Quando faz sentido escolher Track G:

• Você tem genuíno interesse em GPU programming, distributed training, ou modelo internals.
• Mercado tá pagando muito (TC tier-1 em AI Infra > Distributed Systems generic em 2026).
• Você acompanhou 04-10 de perto e quer ir mais fundo do que só consumir APIs.

Trade-offs honestos:

• Stack mexe rápido. Investimento em vLLM 2024 vale 70% em 2026. Re-investment continuous.
• GPU é caro. Side projects sérios = $500-2000/mês em compute. Lambda Labs ou Modal pra burst são alternativas.
• Comunidade ainda fragmentada. Conferences sérias (NeurIPS, ICLR) viraram MLOps/Infra-friendly recente.

Companies: YC startups Series A-B, founding engineer roles, Plasmic, Replit, Linear.

2.11 Cross-track core

Independent of track:

• Logística v1/v2/v3 done.
• Build-from-scratch (2 toys).
• Multi-domain capstones (3).
• Public output stream.
• 3+ mentees grown.
• Paper habit established.

Estes são inputs comum; o capstone cristaliza into portfolio narrative.

2.12 Portfolio assembly

Componentes:

• Personal site (own domain) com about, projects, blog, contact.
• GitHub profile com pinned repos curados.
• LinkedIn alinhado com narrative.
• Talks page (recordings + slides).
• Reading list / now / uses pages opcionais.

Style: professional, specific, evidence-based. Avoid buzzword soup.

2.13 Promo case docs

Two outputs:

• Internal brag doc: para promo em current role. Confidential, with manager.
• External narrative: positioning para next role. Public-friendly version.

Both reference same evidence; different framing.

2.14 Defense ritual

Apresente teu specialization track output a:

• 1+ Senior+ peer (technical critique).
• 1+ manager (strategic critique).
• Optional: external mentor or industry friend.

Iterate baseado em feedback.

2.15 Calibrate vs market

Look at:

• levels.fyi para ranges.
• Job descriptions Staff em empresas alvo.
• Public tech blogs (Engineering Manager Roadmap, Pragmatic Engineer).
• Conf speaker bios.

Compare seu portfolio. Identify gap. Plan.

2.16 Continuação pós-Staff

Capstone fecha framework. Next:

• Operate em Staff role e cresça scope.
• Aspire Principal / Distinguished se shape fit.
• Manager track if appeals.
• Founding engineer if entrepreneurial.
• Sabbatical / open source full-time se financially feasible.

Não há "Stage 6". Trajetórias divergem.

---

3. Threshold de Maestria

Você precisa, sem consultar:

• Justificar specialization vs generalist trajectory pra você.
• Listar componentes de promo case.
• Diferenciar 7 tracks listados em §2.5-§2.11.
• Identificar gap pessoal vs Staff target.
• Aplicar calibration via levels.fyi e job descriptions.
• Estruturar portfolio público.
• Planejar defense ritual.
• Esboçar narrative cross-track (own track vs alternatives).

---

4. Desafio de Engenharia (e Profissional)

Especialização declarada + portfolio + promo case.

Especificação

1. Choose track: doc TRACK-CHOICE.md com justificativa baseada em interesse + market + skills.
2. Capstone showcase:
   • Para track escolhido (§2.5-§2.10), produza outputs específicos completos.
   • 1 deep technical achievement: implementing ambitious thing (Raft real, framework lib, ML pipeline production, etc.).
3. Portfolio site:
   • Próprio domain.
   • About, Projects (com cada output do framework linkado), Blog (de 05-05), Talks, Contact.
   • SEO basics, performance pristine (03-09), accessibility (02-02).
4. Promo case:
   • Internal version (private, com manager).
   • External narrative (public-ish; LinkedIn + portfolio).
   • Evidence index: link a artefatos.
5. Defense ritual:
   • Apresentar a 2+ peers Senior+ via 1h.
   • Coletar feedback estruturado.
   • Iterate.
6. External validation:
   • 1+ talk gravado em conf ou meetup ≥ 50 pessoas.
   • 1+ blog post atingindo > 1k views (proxy de ressonância, não vanity, evidence de público).
   • OSS lib com ≥ 100 stars OU 1+ user empresa documentado.
7. Reading consolidation:
   • 25+ papers lidos (05-04 expansão).
   • Reading list curada e publicada.
8. Cross-track exposure:
   • 1 doc respeitoso explicando quais tracks você não escolheu e por que (mostra clarity, não dismissal).

Restrições

• Não vendetta-document. Promo case factual.
• External narrative respeita NDA / employer.
• Specialization profunda mas não ignora outras (Staff é T-shape).
• Sem fake metrics.

Threshold

• Track declarado.
• Showcase completo.
• Portfolio site live.
• Promo case escrito.
• Defense ritual realizado com feedback documentado.
• 1 talk + 1 broad-reach blog + 1 OSS milestone.

Stretch

• Job mobility test: aplicar a 3 Staff roles em empresas alvo. Mesmo sem mover, signal calibrate.
• Industry contribution: paper, RFC contribution, spec involvement.
• Mentor mentor: tornar mentee de 05-06 que vire Senior ou Staff.
• Side product launched: small but live (track F especially).
• Translate framework pra outra língua (PT-EN se você for monolingual ainda; help others).

---

5. Extensões e Conexões

• Liga com CAPSTONE-fundamentos/apprentice/professional/senior: tudo é fundação.
• Liga com 05-01-05-07: outputs cristalizam aqui.
• Liga com MENTOR.md §8: princípios não-negociáveis aplicam até Staff.

Final do framework. Não há próximo módulo.

---

6. Referências

• "Staff Engineer": Tanya Reilly. Definitivo.
• "The Staff Engineer's Path": Tanya Reilly (livro de mesmo nome).
• "An Elegant Puzzle": Will Larson.
• "Become an Effective Software Engineering Manager": Drozdzal.
• levels.fyi + Pragmatic Engineer newsletter (Gergely Orosz).
• Engineering Ladders public: Rent the Runway, CircleCI, GitLab, Spotify.
• "Slack" book: Tom DeMarco. Career mindset.
• Lara Hogan's "Resilient Management".
• LeadDev conferences (leaddev.com).
• Senior Engineer / Staff Engineer / Principal Engineer talks.
• "How to Become a Top Engineer": Lara Hogan, Camille Fournier essays.
• Engineering blogs of empresas alvo: practice reading their narrative.