AI Engineering Mastery: A 12-Week Production AI Course With Cybersecurity Convergence

What this course is

AI Engineering Mastery is a 12-week premium Applied AI flagship course for software engineers, ML engineers, and security engineers building production AI systems in 2026 and 2027. The curriculum sequences twelve weekly modules across the production AI craft: AI engineering tooling and the 2026 landscape, prompt engineering at production scale, embeddings and vector databases, retrieval-augmented generation patterns, evaluation frameworks, multi-agent architectures, fine-tuning and model selection, multimodal AI applications, production deployment patterns, cost optimization and inference economics, AI safety and responsible deployment, and a capstone production AI system. Cybersecurity convergence is woven throughout because production AI systems cannot ship without addressing prompt injection, data exfiltration, model abuse, and supply chain risk. The course is built on primary sources only: Anthropic, OpenAI, Google DeepMind, and Meta AI official engineering documentation; NIST AI Risk Management Framework (AI 100-1) and the Generative AI Profile (AI 600-1); OWASP LLM Top 10; MITRE ATLAS for adversarial threats; and peer-reviewed academic research from arXiv on retrieval, evaluation, agent design, and inference optimization. Generic AI engineering advice without primary-source backing is excluded. Authored by Julian Calvo, Ed.D. in Learning Sciences with the M.S. Applied AI specializing in Cybersecurity in progress at Northeastern University. Every module pairs reading with a hands-on artifact the learner produces and adds to a production AI portfolio. The capstone is a production AI system shipped against a documented evaluation set with cost, latency, and quality budgets and a documented threat model.

The course follows the dependency order of the production AI craft rather than a textbook chapter order. Week 1 grounds the learner in the 2026 AI engineering landscape and tools so every later module has a tool and a target. Weeks 2 through 11 walk the production AI lifecycle in dependency order: prompts, embeddings, retrieval, evaluation, agents, fine-tuning, multimodal, deployment, cost, safety. Week 12 integrates the work into a production AI system with a documented threat model. Pedagogically the design draws on Kolb's experiential learning cycle (1984) and Bandura's self-efficacy theory (1997): every module sequences a concept, primary-source readings, a hands-on artifact, and a written reflection note. Evidence quality is opinionated. Architecture claims are anchored to AI lab official engineering documentation, peer-reviewed research, or named production case studies disclosed by the operating company. The cybersecurity convergence is anchored to NIST AI RMF, NIST AI 600-1 (Generative AI Profile), OWASP LLM Top 10, and MITRE ATLAS.

What you will learn

• Map any production AI system against the canonical six-layer stack (model, orchestration, retrieval, evaluation, observability, gateway and policy)
• Author production prompts with structured output, prompt caching, evaluation harnesses, and prompt injection defense
• Choose embedding models and vector stores on cost, latency, hybrid search, and twelve-month scale
• Build advanced RAG pipelines with contextual chunking, hybrid retrieval, reranking, citation, and grounded refusal
• Practice eval-first development with capability, robustness, and behavioral evals plus LLM-as-judge calibration
• Compose agent loops with tool budgets, planner-executor and orchestrator-worker patterns, and excessive-agency defenses
• Decide when fine-tuning earns its cost and apply LoRA, QLoRA, and DPO recipes against measured baselines
• Ship multimodal features (vision, voice, document AI) with modality-aware evaluation and prompt-injection defense
• Operate a production model gateway with multi-provider routing, retries, circuit breakers, observability, canaries, and shadow traffic
• Drive cost down with model tier routing, caching, batch inference, quantization, and distillation against measured quality
• Apply NIST AI RMF and OWASP LLM Top 10 controls and run AI red-team passes against the production target
• Ship a capstone production AI system with documented evaluation, threat model, system card, cost model, and rollout plan

12-week curriculum

1. Week 01 · 6h · 4 topics

   The 2026 AI Engineering Landscape and Toolchain

   The Applied AI engineering landscape in 2026, the canonical tool stack (model providers, orchestration, vector stores, evaluation, observability, gateway), the cybersecurity surface every AI engineer owns, and a written tooling decision document the learner returns to in every later module.

   Learning objectives and topics

   Learning objectives.

   • Name the six layers of the production AI stack and the canonical tools at each layer in 2026
   • Map your team's current AI tool stack against the canonical layers and identify gaps
   • Articulate the cybersecurity surface every AI engineer owns: prompt injection, data exfiltration, secret leakage, model abuse, supply chain
   • Pick the model provider, orchestration approach, vector store, and observability layer for a specific production target
   • Produce a one-page AI engineering tooling decision document that anchors every later week's artifact

   Topics.

   • The six layers of the 2026 production AI stack
   • Choosing model providers in 2026
   • The cybersecurity convergence at the AI engineer's desk
   • Authoring the AI engineering tooling decision document

   Assessment: 5 questions · 360 minutes total

2. Week 02 · 6h · 5 topics

   Prompt Engineering at Production Scale

   Production prompt engineering, structured output discipline, prompt caching economics, prompt evaluation harnesses, prompt injection defense, and the prompt portfolio document the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Distinguish exploratory prompting (notebook) from production prompting (versioned, evaluated, observable)
   • Author prompts that produce reliable structured output (JSON schema, tool calls, typed responses)
   • Use prompt caching to drop production cost by 50 to 90 percent on cacheable prefixes
   • Build a prompt evaluation harness that scores prompt versions against a fixed ground truth set
   • Defend against prompt injection at the system prompt, user input, and tool output boundaries

   Topics.

   • Production prompts versus exploratory prompts
   • Structured output discipline: JSON schema and tool calls
   • Prompt caching economics
   • Prompt injection defense at the boundary
   • The week 2 prompt portfolio deliverable

   Assessment: 5 questions · 360 minutes total

3. Week 03 · 6h · 4 topics

   Embeddings and Vector Databases

   Embedding model selection, dimensionality and storage tradeoffs, vector store comparison (pgvector, Qdrant, Pinecone, Weaviate, Turbopuffer), hybrid search (vector plus keyword), and the embedding decision document the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Choose an embedding model on cost, dimensionality, language coverage, and benchmark performance
   • Compare vector stores on cost, latency, hybrid search support, and operational profile
   • Implement hybrid search (vector plus BM25) and explain when hybrid beats pure vector
   • Reason about embedding drift across model versions and the migration cost
   • Author an embedding decision document for the production target

   Topics.

   • Embedding model selection in 2026
   • Vector store comparison
   • Hybrid search: vector plus keyword
   • The week 3 embedding decision document

   Assessment: 5 questions · 360 minutes total

4. Week 04 · 6h · 4 topics

   Retrieval-Augmented Generation Patterns

   RAG architectures (naive, advanced, modular), chunking strategies, query rewriting and HyDE, contextual retrieval, citation and grounding, and the RAG cybersecurity surface (data exfiltration, retrieval poisoning).

   Learning objectives and topics

   Learning objectives.

   • Distinguish naive, advanced, and modular RAG and pick the right pattern for the production target
   • Author chunking strategies that preserve semantic coherence (fixed, semantic, recursive, contextual)
   • Apply query rewriting, HyDE, and multi-query expansion to improve retrieval quality
   • Implement citation and grounding to make RAG outputs verifiable
   • Defend the RAG cybersecurity surface: data exfiltration, retrieval poisoning, sensitive disclosure

   Topics.

   • Naive, advanced, and modular RAG
   • Chunking strategies
   • Query rewriting, HyDE, and multi-query
   • Citation, grounding, and the RAG cybersecurity surface

   Assessment: 5 questions · 360 minutes total

5. Week 05 · 6h · 4 topics

   Evaluation Frameworks for AI Systems

   Eval-first development, ground truth construction, LLM-as-judge patterns, regression evals, online evals (production), evaluation cybersecurity (eval poisoning), and the evaluation harness the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Practice eval-first development: write the eval before the prompt, the chain, or the agent
   • Construct ground truth datasets using human labeling, synthetic generation, and production trace mining
   • Use LLM-as-judge patterns with calibration and bias controls
   • Build a CI evaluation harness that gates prompt and code changes on quality regression
   • Run online evaluations in production and close the loop on quality drift

   Topics.

   • Eval-first development
   • Ground truth construction
   • LLM-as-judge patterns
   • Online evals and evaluation cybersecurity

   Assessment: 5 questions · 360 minutes total

6. Week 06 · 6h · 4 topics

   Multi-Agent Architectures

   Single-agent vs multi-agent design, agent loops, tool use patterns, planner-executor and orchestrator-worker, agent observability, agent cybersecurity (excessive agency, tool abuse), and the agent the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Decide when a multi-agent design beats a single-agent design (and when it does not)
   • Author agent loops with tool use, error handling, and budget enforcement
   • Compose planner-executor and orchestrator-worker patterns for complex tasks
   • Observe agents with traces, replay, and per-step quality scoring
   • Defend the agent cybersecurity surface: excessive agency, tool abuse, prompt injection through tool outputs

   Topics.

   • Single-agent vs multi-agent
   • Agent loops, tool use, and budgets
   • Planner-executor and orchestrator-worker
   • Agent cybersecurity: excessive agency and tool abuse

   Assessment: 5 questions · 360 minutes total

7. Week 07 · 6h · 4 topics

   Fine-Tuning and Model Selection

   When fine-tuning earns its cost, dataset construction, parameter-efficient methods (LoRA, QLoRA), preference tuning (DPO, RLHF), open-weights options (Llama, Mistral), and the model selection decision the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Decide when fine-tuning beats prompting and retrieval (and when it does not)
   • Construct fine-tuning datasets with quality, diversity, and contamination controls
   • Apply parameter-efficient fine-tuning (LoRA, QLoRA) for cost-controlled adaptation
   • Distinguish supervised fine-tuning from preference tuning (DPO, RLHF)
   • Compare hosted fine-tuning APIs against self-hosted open-weights fine-tuning

   Topics.

   • When fine-tuning earns its cost
   • Dataset construction
   • Parameter-efficient fine-tuning: LoRA and QLoRA
   • Preference tuning: DPO and RLHF

   Assessment: 5 questions · 360 minutes total

8. Week 08 · 6h · 4 topics

   Multimodal AI Applications

   Vision (image understanding, OCR, charts), audio (speech-to-text, text-to-speech, voice agents), document AI (PDFs, tables), video, multimodal evaluation, and multimodal cybersecurity (image-borne prompt injection).

   Learning objectives and topics

   Learning objectives.

   • Build vision-enabled features (image understanding, chart and diagram interpretation, OCR)
   • Compose voice agents using speech-to-text, LLM, and text-to-speech with latency budgets
   • Process documents (PDFs, tables, scanned images) with structured output
   • Evaluate multimodal systems against ground truth that respects modality
   • Defend against image-borne and audio-borne prompt injection

   Topics.

   • Vision: image understanding, charts, OCR
   • Voice agents: STT, LLM, TTS with latency budgets
   • Document AI and structured extraction
   • Multimodal cybersecurity: image and audio prompt injection

   Assessment: 5 questions · 360 minutes total

9. Week 09 · 6h · 5 topics

   Production Deployment Patterns

   Gateway architecture, multi-provider routing, streaming, retries and timeouts, rate limiting, observability (traces, metrics, logs), feature flags, canaries, and the deployment runbook the engineer ships.

   Learning objectives and topics

   Learning objectives.

   • Architect a model gateway that handles routing, retries, rate limits, and audit logging
   • Implement streaming end-to-end with backpressure and graceful degradation
   • Compose retries, timeouts, and circuit breakers for AI provider outages
   • Instrument tracing, metrics, and logs across the AI stack with cost and quality dimensions
   • Roll out AI features with feature flags, canaries, and shadow traffic

   Topics.

   • The model gateway pattern
   • Streaming end-to-end
   • Retries, timeouts, and circuit breakers
   • Observability with cost and quality dimensions
   • Feature flags, canaries, and shadow traffic

   Assessment: 5 questions · 360 minutes total

10. Week 10 · 6h · 4 topics

    Cost Optimization and Inference Economics

    The unit economics of AI features, model tier routing, prompt caching at scale, batch inference, quantization, distillation, on-device options, and the cost reduction playbook the engineer ships.

    Learning objectives and topics

    Learning objectives.

    • Compute the unit economics of an AI feature (cost per request, cost per active user, gross margin)
    • Route requests across model tiers to drop cost without dropping measured quality
    • Apply prompt caching, batch inference, and quantization for large cost reductions
    • Use distillation to replace a large model with a small fine-tune at production quality
    • Decide when on-device inference (Llama, Phi, on-device Apple, Gemini Nano) earns its complexity

    Topics.

    • Unit economics of AI features
    • Model tier routing
    • Caching, batching, and quantization
    • Distillation and on-device inference

    Assessment: 5 questions · 360 minutes total

11. Week 11 · 6h · 4 topics

    AI Safety and Responsible Deployment

    NIST AI RMF in practice, OWASP LLM Top 10 application controls, content moderation, jailbreak defense, abuse monitoring, red-teaming, model card and system card discipline, and the responsible deployment runbook the engineer ships.

    Learning objectives and topics

    Learning objectives.

    • Apply NIST AI RMF (Map, Measure, Manage, Govern) to a production AI system
    • Implement OWASP LLM Top 10 controls at the engineering layer (not the policy layer)
    • Build content moderation, jailbreak defense, and abuse monitoring with documented thresholds
    • Run AI red-teaming exercises against the production target
    • Author a system card disclosing scope, limits, and known failure modes

    Topics.

    • NIST AI RMF in practice
    • Content moderation, jailbreak defense, abuse monitoring
    • AI red-teaming
    • Model cards and system cards

    Assessment: 5 questions · 360 minutes total

12. Week 12 · 6h · 4 topics

    Capstone: Ship a Production AI System

    Integrate the eleven prior weeks into a production AI system: scoped task, six-layer architecture, evaluation harness, gateway with policy, observability, cost model, threat model, system card, and rollout plan. The capstone is the work that earns the certificate.

    Learning objectives and topics

    Learning objectives.

    • Integrate the eleven prior weeks' artifacts into a single production AI system
    • Author the architecture, evaluation harness, gateway, observability, cost model, and threat model
    • Author a system card disclosing scope, evaluation results, failure modes, and contact
    • Author a rollout plan with feature flags, canaries, and rollback criteria
    • Pass the three-failure-mode rubric and earn the AI Engineering Mastery certificate

    Topics.

    • Capstone scope and the six-layer architecture
    • Evaluation harness and cost model
    • Threat model, system card, and rollout plan
    • The three-failure-mode rubric

    Assessment: 5 questions · 360 minutes total

Capstone

Ship a production AI system against a documented evaluation, threat model, and rollout plan

The capstone integrates the eleven prior weekly artifacts (tooling decision, prompt portfolio, embedding decision, RAG pipeline, evaluation harness, agent, fine-tune decision, multimodal feature, deployment runbook, cost model, AI safety controls) into a single production AI system. The system has a scoped task, a six-layer architecture, an evaluation harness with capability, robustness, and behavioral layers (50 plus examples each, runs in CI), a gateway with policy, observability with cost and quality dimensions, a documented threat model mapped to OWASP LLM Top 10 and MITRE ATLAS, a system card, a cost model, and a rollout plan with feature flags, canaries, and rollback criteria. The capstone is graded against three named failure modes: no eval harness with ground truth, no threat model addressing prompt injection plus two other LLM Top 10 risks, and no observability with rollback. A passing capstone earns the DecipherU AI Engineering Mastery certificate of completion.

Who it is for

• AI engineers (1 to 5 years building LLM-backed features) wanting deeper production craft
• ML engineers expanding from classical ML into LLM and multimodal production systems
• Software engineers (5+ years) who have shipped LLM features and want to operate the full production AI lifecycle
• Security engineers building AI security tooling who need parity with the AI engineering craft on the other side of the wall
• Platform engineers building AI infrastructure who need to understand the application layer they support
• Tech leads and staff engineers scoping production AI systems at AI-native startups or enterprises adopting AI

Who it is not for

• Engineers without production software experience. Build foundational engineering experience first; this course assumes shipping ownership.
• Engineers who have never called an LLM API. Take AI Career Transition first; this course assumes basic LLM literacy.
• Anyone seeking a vendor exam credential. The course is portfolio-driven and does not certify against any vendor exam.
• Engineers unwilling to ship a production AI system as the capstone. The certificate requires shipped, defendable evidence.
• Engineers seeking guaranteed AI engineering offers at a frontier lab. Hiring outcomes vary by individual and market.

Prerequisites

• At least 3 years of production software engineering experience (shipped, owned, on-call)
• Working familiarity with at least one LLM API (Anthropic, OpenAI, Google) at the prompt and structured output level
• Comfort with Python or TypeScript and at least one production codebase you have shipped end to end
• Basic understanding of vector math (dot product, cosine similarity) and standard CS data structures
• Willingness to commit 75 hours of focused study and artifact production across 12 weeks

What you get

• 75 hours of original Applied AI curriculum across 12 weekly modules with cybersecurity convergence (NIST AI RMF, OWASP LLM Top 10, MITRE ATLAS) woven into every module
• Twelve portfolio-grade artifacts produced across the 12 weeks (tooling decision, prompt portfolio, embedding decision, RAG pipeline, evaluation harness, agent, fine-tune decision, multimodal feature, deployment runbook, cost model, AI safety controls, capstone production AI system)
• Certificate of completion issued for learners who finish all 12 weekly assessments and submit a capstone that passes the three-failure-mode rubric. The certificate is a digital credential with a verifiable URL listing the curriculum, the assessment outcomes, and the capstone artifact (with the engineer's permission).
• Lifetime access to course updates as the production AI landscape evolves (new model versions, new patterns, new cybersecurity findings)
• DecipherU community access (Defender tier and above) for peer review and capstone accountability

Author

Frequently asked questions

Who is this AI engineering cybersecurity-aware course for?

AI engineers (1 to 5 years building LLM-backed features), ML engineers expanding into LLM and multimodal systems, software engineers (5+ years) who have shipped LLM features and want the full production AI lifecycle, security engineers building AI security tooling, platform engineers building AI infrastructure, and tech leads scoping production AI at AI-native or enterprise companies. The course assumes basic LLM literacy and at least 3 years of production software experience.

What primary sources does the course cite?

Anthropic, OpenAI, Google DeepMind, and Meta AI official engineering documentation; NIST AI Risk Management Framework (AI 100-1) and the Generative AI Profile (AI 600-1); OWASP LLM Top 10; MITRE ATLAS; and peer-reviewed academic research from arXiv on retrieval, evaluation, agent design, fine-tuning, and inference optimization. Generic AI engineering advice without primary-source backing is excluded.

How long does AI Engineering Mastery take to complete?

Roughly 75 hours of focused study and artifact production across 12 weekly modules. Most learners complete it in 12 to 16 weeks at 5 to 7 hours per week. Self-paced. The capstone is a production AI system shipped against a documented evaluation, threat model, and rollout plan; it typically takes 2 to 3 weeks of additional work after week 11.

How does this course handle the cybersecurity surface of production AI?

Cybersecurity is woven throughout, not bolted on. Week 1 names the AI engineer's cybersecurity surface across all six stack layers. Week 2 covers prompt injection defense at three boundaries. Week 4 covers the RAG cybersecurity surface (data exfiltration, retrieval poisoning, indirect prompt injection). Week 6 covers excessive agency and tool abuse. Week 11 is the deep dive: NIST AI RMF in practice, OWASP LLM Top 10 controls, content moderation, jailbreak defense, abuse monitoring, AI red-teaming, and system cards. The capstone requires a documented threat model mapped to OWASP LLM Top 10 and MITRE ATLAS.

How is AI Engineering Mastery different from AI Career Transition?

AI Career Transition ($397, 8 weeks) covers the transition arc into AI roles: resume tuning, portfolio shipping, network building, AI engineering interview prep, AI compensation patterns. AI Engineering Mastery (this course, $497, 12 weeks) covers the deeper production AI craft for engineers already inside or adjacent to AI roles: prompt engineering at scale, RAG, evaluation, agents, fine-tuning, multimodal, deployment, cost economics, AI safety, and a shipped production AI system capstone. Pair the two if you need both the transition path and the deeper craft.

Does the course guarantee a job at a frontier AI lab?

No. Hiring outcomes at frontier labs depend on individual fit, market conditions, geography, and many other factors. The course teaches the production AI craft and produces twelve portfolio-grade artifacts plus a shipped capstone the engineer uses in interviews and on-the-job. Hiring panels in 2026 weight portfolio evidence and shipped production AI work heavily; the capstone is the work that produces that evidence.

Companion courses

Need the transition path or the product practice arc instead?

AI Engineering Mastery teaches the deeper production AI craft. AI Career Transition teaches the engineering transition arc into AI roles. AI Product Management teaches scoping AI features, evaluation methodology, and authoring AI product specs that ship. The two companion courses are $397 each and live in the Applied AI foundation catalog.

• AI Career Transition (8-week, $397)
• AI Product Management (8-week, $397)

Related cybersecurity and Applied AI content

• Applied AI vertical overview
• Applied AI career paths catalog
• AI Security Operations Mastery (cybersecurity convergence course)
• Cybersecurity for AI courses (secure AI systems themselves)