AI Engineering · Applied AI Engineer — Learning Repository

🎓 Part of the free, open-source AI Career Curriculum ecosystem — Infrastructure · ML Engineering · AI Engineering · Governance. Live cohorts & team programs: ai-infra-curriculum.github.io.

💜 Sponsor this curriculum — sponsorships keep the whole open-source AI Career Curriculum free and moving.

Ship and operate LLM/agent/RAG features inside real products end-to-end: production prompting, RAG integration, tool calling, product-embedded agents, feature-side evaluation, surface guardrails, model routing, cost / latency SLOs, observability, incident response, and feature-lifecycle ownership.

Status: curriculum plan authored (2026-07-09). Lessons and project READMEs will be drafted by subsequent autonomous content cycles. See CURRICULUM.md for the planned scope and JOB_REQUIREMENTS.md for the requirements-to-coverage map (postings list deferred to the next research cycle — see JOB_REQUIREMENTS.md Status section).

Level: 30 (deep specialist — peer to agentic-ai-engineer and ai-eval-engineer at the same level; one level above the build-altitude peer llm-application-developer at level 25; peer to senior-ml-engineer at the same level on the ML Engineering ladder). Family: AI Engineering. Planned scope: 12 modules (156h) + 3 projects (140h) = ~296 hours.

What this track owns

End-to-end production integration of LLM/agent/RAG features into a product surface:

• Applied AI foundations — spec → LLM-feature architecture, provider / tier / shape choice, SLO mapping, ISO/IEC 25059 & NIST AI RMF framing
• Production prompting & structured outputs — prompt caching, JSON mode / function calling, constrained decoding (Instructor / Outlines), streaming + cancellation, prompt versioning + rollback
• RAG integration at the product surface — consuming a retriever, citations / freshness / evidence UX, retrieval quality instrumentation, query rewriting
• Tool calling & orchestration at product altitude — JSON schemas, MCP servers, LangGraph / OpenAI Agents SDK / Anthropic Claude Agent SDK, per-request budgets
• Shipping product-embedded agents — coding co-pilot, customer-experience, retrieval-with-actions, browsing, sales / support / ops shapes; memory system with retention + PII policy; hand-offs & kill-switches
• Feature-side evaluation — offline suites, snapshot-replay regression fixtures, eval-gated CI, sampled online eval, judge-drift monitor, per-feature dashboards
• Guardrails & safety on the surface — layered moderation, prompt-injection defence, PII scrubbing, tool-abuse guards, OWASP LLM & Agentic Top 10 mapping
• Model routing & provider selection — LiteLLM / OpenRouter / Portkey / Bedrock / Vertex / Foundry gateway patterns, classifier-gated routing, self-host vs managed, multi-modal
• Cost & latency SLOs — TTFT / TPOT at app altitude, prompt caching, batching, per-tenant token / cost accounting, retry / hedging, burn-rate alerts
• Observability, user feedback & incident response — OpenTelemetry Gen-AI / OpenInference instrumentation, sampling + PII scrubbing, Argilla / Label Studio triage, six-failure-mode playbooks, on-call runbooks
• Shipping, experimenting & rollbacks — feature-flag driven rollout (LaunchDarkly / OpenFeature), canary + shadow launches, A/B interface, hot-rollback of a bad prompt, hosted fine-tuning as an applied lever
• Owning an applied-AI feature end-to-end — charter with kill criteria, incident-to-fixture loop, build-vs-buy across gateway / eval / agent SDK, per-feature model-card slice, delegation contracts to peer specialists

What this track defers

• Build-altitude first-time LLM app construction → llm-application-developer-learning (peer, level 25 — the on-ramp to this track).
• RAG engineering depth (chunking, embeddings, vector DB internals, rerankers, hybrid search) → rag-engineer-learning (peer, level 25).
• Agent-framework methodology depth (LangGraph / CrewAI / AutoGen internals, multi-agent orchestration) → agentic-ai-engineer-learning (peer, level 30).
• AI-application evaluation program methodology → ai-eval-engineer-learning (peer, level 30).
• Model-eval methodology depth (validity theory, benchmark construction, statistical rigour, cross-modality methodology, MLPerf serving benchmarks) → model-evaluation-engineer-learning (peer, level 30, ML Engineering family).
• Post-training depth (SFT / PEFT / RLHF / DPO) → fine-tuning-engineer-learning (peer, level 30).
• Alignment-risk methodology, harm modelling, and red-team data generation → ai-risk-engineer-learning (peer, level 30).
• Release-assurance / governance shape (audit trails, regulator-facing docs, third-party evaluator interface) → ai-evaluation-engineer-learning (peer, level 25, Governance family).
• MLOps / model registry / self-service ML/AI platform / inference serving → ai-infra-mlops-learning, ai-infra-ml-platform-learning, ai-infra-performance-learning.
• Classical ML fundamentals (PyTorch, sklearn, FastAPI packaging, MLflow) → ml-engineer-learning (level 20).
• Deep ML/AI security (model extraction, eval-set exfiltration, prompt-injection defence at platform depth) → ai-infra-security-learning.
• Governance / policy / model-card review depth → ai-governance-analyst-learning, head-of-ai-governance-learning.

Layout

applied-ai-engineer-learning/
├── .aicg/                    curriculum-plan.json and job-requirements.json (machine-readable catalog)
├── lessons/mod-XXX-*/        modules with lectures, exercises, labs, quizzes (to be drafted)
├── projects/project-XXX-*/   multi-module capstones (to be drafted)
├── CURRICULUM.md             role-level coverage map
├── JOB_REQUIREMENTS.md       requirements catalog with citations and ownership map
├── PREREQUISITES.md          assumed entry skills
├── VERSIONS.md               release history
└── README.md                 this file

Paired Solutions Repo

applied-ai-engineer-solutions carries the reference implementations.

---

Maintained by VeriSwarm.ai