AI Engineer — Role-Based Learning Hub
A modular collection of role-specific, project-driven curriculum tracks. Each track targets a distinct job description, stands alone as a complete learning path, and opens as its own book with a focused sidebar and search.
Published at https://ai-engineer-zt0.pages.dev/ — built from master on every push.
Tracks
| Track | Focus |
|---|---|
| LLM Inside-Out Mastery | Vocabulary → transformers → model selection → local inference → serving → RAG → agents → eval → security → startup |
| LLM Inference Engineer | Tokenization → transformers → training → RAG → inference serving |
| CV Engineer | Classical CV, PyTorch/TF, detection & segmentation, MLOps |
| Senior ML Engineer | Feature stores, retrieval, ranking, recsys, experimentation, MLOps |
| Model Accuracy & AI Performance | Framework internals, quantization, compilers, NPU deployment, profiling |
| Head of Software Engineering — GPU | GPU architecture, CUDA, compilers, drivers, runtimes, scheduling |
| Rack Management — Senior/Staff SWE | Rack-scale AI systems, Redfish/IPMI/SNMP, BMC/PDU/CDU/PCIe, provisioning, K8s, telemetry, firmware/RAS, secure multi-tenancy |
| AI Specialist | Operational digital twin (JD1) + inference optimization & accelerators (JD2) |
| Apache Committer / PMC Engineer | ASF governance, JVM systems, review culture, release engineering |
| Security Engineer | AppSec, OS & sandboxing, cloud, vuln research, IR, AI/agent security |
| Principal Data Infrastructure Engineer | Distributed data platform: Kafka/Kinesis, Flink, Spark, EMR, Iceberg/Delta/Hudi, Trino/Athena, Cassandra, Scala/Cats/ZIO, governance & reliability |
| Red Team Engineer | Adversary emulation, engagement lifecycle, OS internals, AD/Kerberos, C2, cloud/container, social engineering, RE, purple team — Mandiant consulting track |
| Principal Azure Cloud Engineer & Architect | ARM control plane, Terraform/CDKTF IaC, Entra ID/OAuth2/OIDC/JWT, RBAC & Policy, landing zones, VNet/NSG/routing, ACR/AKS, CI/CD + OIDC federation, APIM, Service Bus/Event Grid, Functions/Durable, Key Vault/Managed Identity, KQL, reliability & FinOps |
| Senior AI Engineer (LLMs/VLMs/Agents) | From scratch: tokenizer → transformer/attention → autograd → VLMs → LoRA/QLoRA/quantization → RLHF/DPO → sampling & constrained decoding → vLLM-class serving (PagedAttention) → distributed training → RAG/ANN → agents & multi-agent → neurosymbolic → embodied/VLA → eval/guardrails → MLOps (MLflow) → CUDA → capstone |
| Agentic AI Engineer (Infra & Platforms) | Agent infrastructure from scratch: reliability/cost math → ReAct/ReWOO loop → tool calling & JSON-Schema → MCP servers → context engineering → RAG + GraphRAG/LightRAG/RAPTOR → multi-agent → durable execution (Temporal-class) → sandboxing → prompt-injection defense → LLM-as-judge evals → async services → multi-tenancy → cost/observability → coding agents → voice → enterprise-platform capstone. Part II — Frameworks Deep Dive: LangGraph, CrewAI, AWS Bedrock AgentCore, OpenAI Agents SDK, Google ADK, AutoGen/Microsoft Agent Framework, and the Amazon Bedrock foundation-model platform. Part III — Cloud AI Platforms, MLOps & Production Infra: SageMaker/Vertex AI/Azure ML, MLOps (tracking/registry/CI-CD/drift), data & feature engineering (feature stores/Dataflow/BigQuery ML/Databricks), Kubernetes/OKE SRE. Part IV — GenAI Frameworks & ML Foundations: LangChain core, Hugging Face, Cohere, ML/DL foundations (TF/PyTorch, supervised/unsupervised/RL). Part V — Eval-Driven Development for software-delivery agents. Every phase is a faithful stdlib miniature + four principal-depth deep-dive docs (Deep Dive / Principal Deep Dive / Core Contributor Notes / Staff Engineer Notes) |
| Enterprise AI & Agentic Platform — Engineering & Architecture | The senior technical authority for a Tier-1 bank's AI & Agentic Platform, two-in-a-box with the Platform Product Owner, under CBUAE regulation. Builds the substrate a whole bank's agents run on: the five-layer stack (Action Gateway, Agent Kernel, Control Plane, Knowledge Foundation, Users & Channels) plus the cross-cutting model, identity and infrastructure layers. Platform availability & error-budget math → the agent kernel (lifecycle, tiered memory, externalized state, consistent-hash session affinity) → the multi-protocol stack (MCP tool estate with versioning and authorization-aware discovery; A2A/ACP delegation with depth, cycle, tenancy and classification admission) → the LLM gateway (routing, fallback, semantic caching, token accounting, tenant isolation) → serving economics (PTUs vs PAYG, vLLM/KV-cache, batching) → knowledge foundation (hybrid retrieval) and FIBO/OWL/SHACL/SPARQL financial knowledge graphs → agent & workload identity (NHI, OAuth 2.1, RFC 8693 token exchange, SPIFFE, mTLS, JIT credentials) → zero trust, KYA and policy-as-code → the action gateway (idempotency, sagas, breakers, hash-chained audit) → runtime guardrails (PII/MNPI, injection defence, HITL, OWASP LLM Top 10) → integration fabric (ISO 20022, Kafka, outbox, data products) → cloud backbone (Terraform, AKS, mesh, private networking) → SRE for non-deterministic workloads (SLOs, burn-rate alerting, OTel GenAI, FinOps) → CBUAE/model-risk evidence packs → two-in-a-box operating model → a bank-grade capstone. Every phase is a runnable stdlib miniature plus six teaching docs |
| Elite AI-Lab SWE Interview Program | Not a track — an executable interview-prep program for OpenAI/Anthropic/DeepMind-class SWE loops. Phase-0 research with confirmed/reported/inference labelling → a timed 3-hour baseline diagnostic that assigns a level per track → a 26-week plan built from the measurement. Seven tracks: a stage-gated progressive coding harness (15 problems, gate N+1 hidden until gate N passes — versioned KV store, token-stream differ with rollback), Python internals proven by runnable experiments (generators, async/TaskGroup, the GIL race that modern CPython hides, __slots__/memoryview measured), distributed design with back-of-envelope calculators and 12 critiqued design artifacts, inference infra (KV-cache math, roofline, continuous batching, autoscaling on token/queue signals) with gpu_math.py, a 48-hour take-home rehearsal plus a deep-dive interrogation harness that reads your actual diff, staff-altitude behavioral, and agent-driving. Weekly scored mocks on a no-hire → strong-hire-at-staff scale, spaced repetition at 1/3/7/21 days |
| Systems From First Principles | Not a curriculum — a 34-month, 15-project build-and-measure journey that trains you to reconstruct foundational systems from their constraints, then measure, break, modify and publish them. Transformer (RoPE, cost model, KV cache) → ANN index (HNSW, recall/QPS, the two-factor speedup model) → vector DB (WAL, filtering cliff, compaction) → LSM engine (Bloom theory vs measurement, the three amplifications) → distributed KV (Raft, replayable fault injection, linearizability checking) → MapReduce (stragglers, speculative execution) → stream processor (watermarks, exactly-once effects) → recommender + simulator + A/B platform (popularity bias, offline-vs-online divergence) → language and bytecode VM with GC → OS kernel (syscall/context-switch cost, Bélády's anomaly) → tensor framework and autodiff → hardware-aware ML (tiling, roofline, a systolic-array simulator) → an integrated system answering one research question, written up as a paper. Every project runs the same 14-step loop: naive design before reading the paper, predicted failure, instrumentation, baseline, falsifiable hypothesis, controlled experiment, failure analysis, report. Includes five executed tools (bench.py, roofline.py, metrics.py, annlab.py, bloom.py), a worked research-notebook entry with real measured numbers, a 12-category scorecard, and a strict AI-assistant policy |
| Frontier Pre-Training Lead | The job Google DeepMind's Gemini pre-training lead actually does: turn a compute budget into the best servable model. FLOPs/memory algebra (C = 6ND) → scaling laws (Kaplan → Chinchilla → IsoFLOPs, robust fitting, bootstrap forecasts, optimal ladder design) → Mixture of Experts from scratch (routing, load-balance & z-loss, capacity/dropping, total-vs-active, the all-to-all wall). Includes a full dissection of Vlad Feinberg's interview and his Gemini Pretraining talk — MoE, MFU, distillation, quantization, pipelined prefill and the Flash 2.0 war story, explained from zero background with runnable code |
How to Use This Hub
Search across every book from the box at the top of this page — press
s or / to jump to it, narrow to one track with the dropdown,
and share a result set with the #q= link that appears in the address bar. Inside
a track, that track's own sidebar search is the faster tool.
- Pick the track that matches your target role and open its book.
- Start with the track's Overview for the full roadmap and weekly schedule.
- Work through phases sequentially — each phase gates the next.
- Use each track's Interview Prep and System Design chapters as running references.
- Use the ← All Roles link at the top of any track's overview to come back here.
Cross-Track Skills (Shared Foundations)
Regardless of which track you pursue, these skills underpin every role:
- Python & systems fluency — typing, packaging, profiling, concurrency.
- Linear algebra, probability, and optimization — the math beneath every model.
- PyTorch internals — autograd, modules, data pipelines, distributed basics.
- Production engineering — testing, CI, containers, observability.
- System design — every track ends in design walkthroughs for its domain.
Published · commit
ec521c1· 2026-08-08 20:32 UTC