🧠 Roadmap AI Engineering Stack — Bare-Metal On-Prem & Homelab

Filosofi

Jangan hanya “fine-tune di Colab” — kamu harus bisa “operate.” Bedanya besar: menjalankan ollama run llama3 itu 2 menit, tapi operate multi-model gateway dengan KV cache optimization, RAG reranking, hallucination tracking, dan automated red-team pipeline itu skill yang membedakan engineer dari notebook hobbyist. Rekruter akan tanya: “Latency melonjak 400ms saat context window penuh, apa yang terjadi di backend dan bagaimana kamu mitigasi tanpa ganti hardware?” Kalau kamu jawab: “KV cache eviction policy diaktifkan, speculative decoding di-disable untuk prioritas throughput, fallback ke smaller quantized model via LiteLLM router, dan Langfuse trace menunjukkan bottleneck di embedding retrieval, bukan LLM inference” — itu yang menutup pertanyaan.

🎯 Checkpoint Awal — Sebelum Mulai

  • Stack: Ollama / vLLM (single-node) → FastAPI → PostgreSQL + pgvector / Qdrant
  • Jalur: AI Engineer → MLOps/LLMOps Engineer → AI Infrastructure Lead
  • Spek Dasar: i7 Gen7, 8GB RAM, GTX 1050 (⚠️ VRAM 2GB sangat terbatas. Kuantisasi 4-bit/8-bit, CPU offload, atau cloud GPU spot instance wajib untuk fase 3+)
  • Target Karir: AI Systems Engineer → AI Platform Lead → MLOps Architect

Homelab Strategy untuk VRAM/RAM Terbatas

  • Fase 1-2: CPU inference + Q4_K_M quantization, model ≤ 3B, Docker resource limits ketat.
  • Fase 3-4: RAG pipeline jalankan terpisah dari inference server, gunakan SQLite/Chroma dulu sebelum Qdrant untuk hemat RAM.
  • Fase 5-6: Cloud GPU trial (RunPod / Vast.ai ~$0.20/jam) hanya untuk evaluation benchmark & red-team stress test.
  • Prioritas: Data pipeline reproducibility → Serving latency control → Evaluation metrics → Security guardrails.

Urutan Belajar

  1. Fase 1 (Foundation): Local serving + quantization + OpenAI-compatible API
  2. Fase 2 (Data): Ingestion pipeline + chunking strategy + vector DB + embedding selection
  3. Fase 3 (Architecture): Advanced RAG + reranking + agent state machine + caching
  4. Fase 4 (MLOps): Experiment tracking + DVC + CI/CD for AI + versioned prompts
  5. Fase 5 (Observe & Secure): Tracing + evaluation framework + guardrails + automated red team
  6. Fase 6 (Operate & Scale): Multi-model routing + inference optimization + AI compliance + cost control

Next step: Install Ollama/vLLM, pull Qwen2.5-1.5B-Q4_K_M, expose OpenAI-compatible API via FastAPI, ukur tokens/sec dan TTFT (Time To First Token).

Aturan Emas

Satu pipeline dulu. Selesaikan sampai bisa reproduce experiment, track version, monitor latency, dan block malicious prompt secara otomatis, baru loncat ke layer berikutnya. AI infrastructure adalah rantai probabilistik — jika satu link lemah (data, cache, guardrail), seluruh output jadi tidak trustworthy.


Fase 1 — Local Model Serving & Quantization (Minggu 1–4)

Goal: Kuasai inference engine, quantization trade-off, dan API standardization. Resource Impact: vLLM/Ollama ~2GB RAM + VRAM tergantung model. Q4_K_M 3B model ≈ 2GB VRAM aman.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
Ollama / vLLMHost ~1–2GB RAM + VRAMPagedAttention, KV cache management, quantization formats (GGUF/AWQ), continuous batching, OpenAI API compatibilityvLLM + benchmarking script = kamu bisa ukur throughput (tokens/sec) vs context length secara empiris
FastAPI~256MBAsync routing, request validation (Pydantic), streaming response (SSE), health check, rate limiting middlewareFastAPI + vLLM backend = kamu punya production-ready inference endpoint dalam <50 line kode
Docker / Compose~512MBMulti-stage build, resource limits (CPU/RAM/VRAM passthrough), network isolation, volume management untuk model weightsDocker + reproducible inference = setup berjalan identik di laptop, server, atau cloud tanpa dependency hell
PyTorch Basics~1GBTensor shape, device placement, model loading, torch.no_grad(), memory profiling (torch.cuda.memory_summary)PyTorch + memory audit = kamu paham kenapa OOM terjadi dan cara debug tanpa tebak-tebakan

Cara Belajar Fase 1

Deploy vLLM di bare metal. Expose OpenAI-compatible API. Kirim 100 concurrent request. Ukur TTFT (Time To First Token), P99 latency, dan GPU utilization. Ganti quantization Q4 → Q8 → FP16. Dokumentasikan trade-off. Ulangi sampai pipeline tidak crash saat context window penuh.

Proyek Portofolio Fase 1: Bare-Metal LLM Inference Server — dokumentasi lengkap: vLLM config, quantization benchmark table, Docker Compose file, FastAPI streaming endpoint, load test result (k6/Locust) dengan grafik latency vs concurrency.


Fase 2 — Data Pipeline & Vector Infrastructure (Minggu 5–8)

Goal: Bangun ingestion pipeline yang reproducible, kuasai chunking strategy, dan implement vector search. Resource Impact: Embedding model ~1GB VRAM, Vector DB ~512MB–1GB RAM, ETL tools ~256MB.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
MinIO / Local Storage~256MBS3-compatible API, versioning, lifecycle policy, TLS for data at restMinIO + structured dataset version = raw data tidak pernah hilang atau teroverwrite tanpa audit trail
pgvector / Qdrant~512MBHNSW index, cosine vs dot product, metadata filtering, ANN vs exact search trade-offpgvector + recall@K metric = kamu bisa justify kenapa index tertentu dipilih berdasarkan query pattern
LangChain / LlamaIndex Core~256MBDocument loaders, text splitters (recursive, semantic, markdown), embedding pipeline, vector store integrationLlamaIndex + chunking benchmark = kamu bisa tunjukkan recall improvement dari 300-token fixed vs semantic chunking
Unstructured / DuckDB~512MBOCR pipeline, table extraction, schema validation, SQL-on-JSON/ParquetUnstructured + DuckDB ETL = kamu bisa transform PDF/HTML menjadi structured records tanpa regex hack

Realita Chunking

Ukuran chunk bukan parameter ajaib. Strategi chunking menentukan recall lebih besar daripada model size. Chunk terlalu kecil = context hilang. Chunk terlalu besar = noise masuk. Selalu ukur recall@K dan precision sebelum deploy.

Proyek Portofolio Fase 2: Document Ingestion & Vector Search Pipeline — diagram data flow, chunking strategy comparison table, embedding model benchmark (speed vs quality), pgvector index config, search API dengan recall/precision metrics.


Fase 3 — Advanced RAG & Agent Architecture (Minggu 9–12)

Goal: Implement hybrid search, reranking, query rewriting, dan agent loop dengan state machine yang deterministic. Resource Impact: Reranker model ~1GB VRAM, Cache/Agent ~512MB RAM, Redis ~256MB.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
Advanced RAG Patterns~256MBHybrid search (BM25 + vector), query expansion, self-query, context compression, reranking (Cross-Encoder)Reranker + precision@3 metric = kamu bisa tunjukkan peningkatan relevansi tanpa tambah context window
LangGraph / Custom State Machine~512MBFinite state machine for agents, tool routing, human-in-the-loop, checkpoint/rollback, deterministic vs stochastic flowLangGraph + tool failure recovery = agent tidak hang atau loop tak terbatas saat tool error
Redis / Memcached~256MBSemantic cache, TTL management, cache invalidation strategy, hit-rate monitoringRedis + semantic cache = kamu reduce API call 40% tanpa degrade response quality
Nginx / HAProxy~128MBAPI routing, request buffering, timeout management, circuit breaker patternHAProxy + fallback routing = traffic otomatis switch ke smaller model saat primary latency > threshold

Agent Tanpa State Machine = Chaos

Framework agent yang mengandalkan “LLM memutuskan sendiri kapan stop” tidak production-ready. Selalu gunakan explicit state graph, define retry limits, dan instrument setiap transition. Determinism > autonomy di production.

Proyek Portofolio Fase 3: Production RAG Agent — architecture diagram (retriever → reranker → generator → tool executor), cache hit-rate dashboard, state machine config, latency breakdown per component, failure recovery log.


Fase 4 — MLOps, Experiment Tracking & CI/CD for AI (Minggu 13–16)

Goal: Version control data, model, dan prompt. Otomasi evaluasi dan deployment pipeline. Resource Impact: MLflow/DVC ~512MB, CI Runner ~1GB, Monitoring ~256MB.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
MLflow / Weights & Biases~512MBExperiment tracking, model registry, artifact versioning, parameter logging, comparison UIMLflow + reproducible run = kamu bisa rollback ke experiment spesifik dalam 1 command
DVC (Data Version Control)~256MBData pipeline tracking, remote storage sync, pipeline DAG, metric loggingDVC + Git = data dan kode selalu sinkron, tidak ada “model ini pakai data yang mana?”
GitHub ActionsN/AAI-specific CI: lint prompts, run evaluation suite, build Docker, push to registry, deploy to stagingGitHub Actions + automated eval = model tidak deploy jika hallucination rate > 5%
Evidently AI / Prometheus~512MBData drift detection, concept drift, feature distribution shift, alerting integrationEvidently + drift alert = kamu detect degradation sebelum user komplain

Prompt Adalah Kode

Jika prompt tidak di-version control, di-review, dan di-test, itu technical debt yang menunggu waktu. Treat prompt seperti production config. Simpan di Git, review via PR, test via automated eval.

Proyek Portofolio Fase 4: AI Model & Prompt CI/CD Pipeline — repo structure, DVC pipeline YAML, MLflow experiment dashboard screenshot, GitHub Actions workflow (eval → gate → deploy), prompt versioning history, rollback procedure.


Fase 5 — Observability, Evaluation & Security (Minggu 17–20)

Goal: Ukur hallucination, latency, cost, dan implement guardrail + automated red team. Resource Impact: Langfuse/Otel ~512MB, Garak/PyRIT ~1GB, Guardrails ~512MB.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
Langfuse / OpenLLMetry~512MBDistributed tracing, span annotation, cost tracking, user feedback loop, session replayLangfuse + trace correlation = kamu bisa debug hallucination dari raw prompt sampai final token
DeepEval / Ragas~256MBAutomated evaluation metrics: faithfulness, answer relevancy, context precision, hallucination scoreRagas + CI gate = deploy diblokir otomatis jika faithfulness < 85%
Garak / PyRIT / LlamaGuard~1GBAutomated red teaming, prompt injection testing, safety classification, guardrail integrationGarak + LlamaGuard = kamu punya pre-deploy security audit report yang terdokumentasi
NeMo Guardrails~512MBInput/output flow definition, regex + LLM-based filters, dynamic policy update, fallback routingNeMo + Colang flows = prompt injection ter-block tanpa matiin seluruh endpoint

Evaluasi Tanpa Baseline = Teater

Jangan percaya angka “accuracy 95%” tanpa tahu baseline, dataset composition, dan edge cases. Selalu bandingkan dengan deterministic baseline (keyword search, rule-based system) dan dokumentasikan failure mode.

Proyek Portofolio Fase 5: AI Observability & Security Dashboard — Langfuse trace example, Ragas evaluation report, Garak scan output, NeMo guardrail config, hallucination trend chart, incident response playbook untuk AI failure.


Fase 6 — Advanced Optimization, Scale & Compliance (Minggu 21–24)

Goal: Model routing, inference optimization, AI governance, dan cost-aware operation. Resource Impact: Router/Proxy ~256MB, Optimization tools ~512MB, Compliance docs ~N/A.

Tool/StackResourceYang DipelajariCombo A+B yang Membuktikan
LiteLLM / OpenRouter Proxy~256MBMulti-model routing, cost/latency balancing, fallback chains, API key rotation, rate limit distributionLiteLLM + cost tracker = kamu route request ke model termurah yang masih memenuhi SLA
vLLM Advanced / Speculative Decoding~1GBDraft-verify pipeline, KV cache optimization, continuous batching tuning, tensor parallelism konsepSpeculative decoding + throughput chart = latency turun 30% tanpa akurasi hilang
AI Compliance & AuditN/AEU AI Act mapping, NIST AI RMF, data privacy (GDPR/PII redaction), model card documentation, bias auditCompliance pack + automated PII check = kamu siap audit tanpa manual gathering
Cost & Resource Management~256MBGPU utilization monitoring, auto-scaling logic, spot instance strategy, egress cost control, cache hit ROIPrometheus + cost dashboard = kamu bisa justify infra spend berdasarkan ROI per request

Proyek Portofolio Fase 6: Multi-Model AI Gateway + Compliance Pack — routing config, speculative decoding benchmark, EU AI Act control mapping table, PII redaction test result, cost/latency optimization report, audit trail sample.


🗺️ Roadmap Visual — Timeline 6 Bulan

Bulan 1Bulan 2Bulan 3Bulan 4Bulan 5Bulan 6
FASE 1
Serving & Quant
FASE 2
Data & Vector
FASE 3
RAG & Agent
FASE 4
MLOps & CI/CD
FASE 5
Observe & Secure
FASE 6
Scale & Comply
vLLMMinIOLangGraphMLflowLangfuseLiteLLM
FastAPIpgvectorRerankerDVCDeepEvalSpeculative Dec
DockerLlamaIndexRedis CacheGitHub ActionsGarakAI Compliance
PyTorchUnstructuredHAProxyEvidentlyNeMo GuardCost Tracking
BenchmarkChunking StrategyState MachinePrompt VersioningRed Team ReportRouting Policy
Portfolio 1:
Inference Server
+ Load Test
Portfolio 2:
Ingestion Pipeline
+ Vector Search
Portfolio 3:
RAG Agent
+ Cache + State
Portfolio 4:
AI CI/CD
+ DVC + MLflow
Portfolio 5:
Observe Dashboard
+ Guardrail
Portfolio 6:
Multi-Model Gateway
+ Compliance

🎓 Sertifikasi yang Cocok per Fase

FaseSertifikasi / MilestoneKenapa
Setelah Fase 1NVIDIA DLI: Fundamentals of Accelerated ComputingValidasi pemahaman GPU memory management & inference basics
Setelah Fase 2Databricks: Data Engineering AssociatePipeline reproducibility, ETL best practice, vector DB foundation
Setelah Fase 3Linux Foundation: MLOps FundamentalsAgent architecture, caching, routing, production patterns
Setelah Fase 4MLflow Certified Practitioner (atau portfolio setara)Experiment tracking, model registry, CI/CD for AI adalah standar industri
Setelah Fase 5OWASP AI Security & MITRE ATLAS AwarenessGuardrail, red teaming, observability, hallucination tracking
Setelah Fase 6NIST AI RMF Compliance Mapping (self-audit)Governance, cost control, multi-model routing, audit readiness
Jangka PanjangKubernetes AI/ML Operator / Custom PlatformScale-out, distributed inference, enterprise AI platform architecture

⛔ Yang TIDAK Perlu Dipelajari Sekarang

Jangan Buang Waktu

  • Cloud-managed AI penuh (SageMaker, Vertex AI, Azure ML) — fokus stack ini on-prem/bare-metal, kontrol penuh > kemudahan terkelola.
  • Training model dari nol (pre-training/fine-tuning besar) — ini track Research/ML Engineer, bukan AI/Platform Engineer. Fokus pada serving, RAG, MLOps, security.
  • Frontend/UI framework (React, Streamlit, Gradio) — kamu build backend pipeline & infrastructure, bukan dashboard. UI adalah konsumen, bukan core stack.
  • Chasing SOTA paper tanpa deployment context — paper penting untuk referensi, tapi production mengutamakan latency, cost, reproducibility, dan reliability.
  • Prompt engineering kursus “rahasia” — prompt adalah config, bukan skill mistis. Pelajari versioning, testing, dan guardrail.
  • Entry-level data cleaning / manual labeling — otomasi ingestion, chunking, dan evaluation. Engineer tidak label data manual di production.
  • CISSP / Compliance umum tanpa AI context — fokus pada EU AI Act, NIST AI RMF, dan data privacy spesifik untuk LLM/AI.

🔗 Lihat Juga

Roadmap AI Engineering | Fase 1 (Serving) → Fase 6 (Scale & Compliance) · 6 Bulan Homelab · Bare-Metal On-Prem Focus