ittwist logo

Roadmap

The AI engineer roadmap

For software developers moving into AI engineering. Follow the 10 Core stages in order for the shortest complete path; add only the 5 Specializations your role needs. Pick one primary resource per stage and keep evolving one real project.

Core 0/218 · All 0/301 · 0%

Start here

Prerequisites

You should have built one small program and know basic Python, Git, SQL, HTTP, and the shell. Missing one? Use Stage 01 to close the gap. No prior ML experience is required.

One project rule

Build one vertical slice

Start a narrow app in Stage 03. Add retrieval or tools, then evals, security, tracing, deployment, and user evidence to that same app. Resource projects are practice, not extra portfolio requirements.

Job-ready checkpoints

After Stage 03

AI prototype ready

Ship a structured-output app with a fixed test set and a measured latency and cost baseline.

After Stage 07

Production feature ready

Ship a grounded RAG or agent workflow with regression evals, tracing, and a threat model.

After Stage 15

Job-ready portfolio

Deploy the capstone with an SLO, rollback path, design document, public demo, and honest postmortem.

Choose for your role

Specialization tracks

Application & agents

Agents, Data & Memory, Multimodal, Product UX

RAG & search

RAG, Data & Memory, Adaptation, Evaluation

Models

Adaptation, Inference & Hardware, Multimodal

AI platform

Inference & Hardware, Production, Data & Memory, Evaluation

01

Software Engineering Foundations

Build maintainable services first; AI systems inherit every ordinary software failure mode.

CoreEffort L0/19
Done when: Expose one typed API with SQL persistence; CI must run unit and integration checks and reject lint or type errors.

Curated resources

Reference · PythonThe Python TutorialThe canonical language foundation.Course · MITThe Missing SemesterShell, Git, editors, debugging and practical tooling.Read · GoogleEngineering PracticesPractical standards for reviews, readability and change quality.Read · Cosmic PythonArchitecture Patterns with PythonBoundaries, testing and maintainable service design.Build · CodeCraftersBuild Your Own XPractice by rebuilding one small tool you use daily.

02

ML & Deep Learning Literacy

A high-level refresher: understand model behaviour and trade-offs, not how to train foundation models.

CoreEffort S0/15
Done when: Train and evaluate one baseline on held-out data; report the metric and document one leakage or imbalance check.

Curated resources

Course · GoogleMachine Learning Crash CourseUse selected modules as a concise refresher.Watch · 3Blue1BrownNeural NetworksVisual intuition without a research syllabus.Build · KaggleKaggle LearnPractice building and measuring one honest baseline.Course · DeepLearning.AIMachine Learning SpecializationOptional if your ML foundations are genuinely missing.Course · fast.aiPractical Deep Learning for CodersOptional for roles that train or fine-tune models.

03

LLM Foundations

Understand how LLMs behave in applications, where they fail, and which lever to use.

CoreEffort M0/22
Done when: Add structured output or a tool call to the project; reach 95% schema-valid output across 50 fixed cases and publish a cost-per-request model.

Curated resources

Course · DeepLearning.AIGenerative AI with LLMsA practical overview of the LLM lifecycle.Read · Jay AlammarThe Illustrated TransformerThe clearest visual transformer explanation.Watch · Andrej KarpathyIntro to Large Language ModelsA one-hour working model for builders.Build · Andrej KarpathynanoGPTOptional practice: train a tiny GPT once so the abstraction stops being magic.Course · Hugging FaceThe LLM CourseOptional hands-on depth for open-source models.Reference · AnthropicPricingRead a real price sheet: input, output, cached and batch rates per tier.Reference · AnthropicPrompt CachingThe largest single cost lever in most applications.

04

Retrieval-Augmented Generation

Build retrieval that is grounded, measurable, and better than simply adding more context.

CoreEffort L0/26
Done when: Add cited retrieval to the project; evaluate at least 50 queries and report recall@5 and groundedness before and after tuning.

Curated resources

Course · Stanford CS276Information Retrieval and Web SearchThe durable theory behind search and ranking.Course · DeepLearning.AIBuilding and Evaluating Advanced RAGAdvanced retrieval patterns with evaluation.Read · PineconeRetrieval-Augmented GenerationA practical, system-level introduction.Reference · PostgreSQL communitypgvectorLearn vector search without adding another database.Build · Exploding GradientsRagasUse it to score your project index before tuning anything.

05

Agents & Context Engineering

Use workflows first, then add autonomy only where the task truly needs it.

CoreEffort M0/24
Done when: Add one checkpointed workflow with retry and approval; 20 adversarial cases must stop within its declared step and cost budgets.

Curated resources

Read · AnthropicBuilding Effective AgentsThe clearest guide to workflows, agents and patterns.Read · AnthropicEffective Context EngineeringTreat context as a scarce engineering resource.Read · AnthropicWriting Tools for AgentsTool design is where most agent quality is won or lost.Course · BerkeleyAdvanced LLM Agents MOOCA research-to-production agents curriculum.Reference · MCPModel Context ProtocolThe official protocol concepts and implementation path.Reference · A2AAgent2Agent ProtocolOfficial interoperable agent communication standard.Reference · AG-UIAG-UI ProtocolThe third protocol: MCP for tools, A2A for agents, AG-UI for the user.Build · LangChainLangGraphUse it if your project needs a durable, checkpointed graph.

06

Evaluation & Observability

Define quality, find failure modes, and prevent regressions before adding more features.

CoreEffort M0/24
Done when: Create at least 50 golden cases; CI must catch one seeded regression and traces must expose latency, cost, and tool calls.

Curated resources

Read · Hamel HusainYour AI Product Needs EvalsA practical method for domain-specific eval systems.Reference · OpenAIWorking with EvalsDataset, grader and evaluation workflow fundamentals.Course · DeepLearning.AIEvaluating AI AgentsTrace-level evaluation for agent systems.Reference · OpenTelemetryGenAI Semantic ConventionsPortable telemetry for models and agents.Build · LangfuseLangfuseTrace your project end to end and wire its regression suite.

07

Security, Safety & Governance

Assume every prompt, retrieved document, model output, and tool call can be hostile.

CoreEffort S0/24
Done when: Threat-model the project and test 20 prompt-injection cases; block unauthorized data and tool access and document residual risk.

Curated resources

Reference · OWASPOWASP Top 10 for LLM ApplicationsThe essential application-security checklist.Read · NIST AI RMFGenerative AI ProfileCross-sector governance and risk controls.Reference · MITREATLASReal adversary tactics and mitigations for AI systems.Build · NVIDIAgarakScan your project and triage what it finds.

08

Data, Storage & Memory

Design the data lifecycle behind retrieval, agents, and durable user experiences.

SpecializationEffort M0/20
Done when: Add an idempotent ingestion pipeline with deduplication, provenance, retries, and a verified delete path for one user or tenant.

Curated resources

Reference · LangGraphMemory in LangGraphA concrete short- and long-term memory model.Read · UC BerkeleyMemGPTThe influential virtual-context memory paper.Reference · PostgreSQLPostgreSQL TutorialA durable base for application and vector data.Build · DataTalks.ClubData Engineering ZoomcampUse selected modules to strengthen your project's ingestion pipeline.

09

Multimodal: Vision, Documents & Voice

Most real workloads arrive as PDFs, screenshots, and speech — not clean text.

SpecializationEffort M0/13
Done when: Process 20 representative files and meet a declared extraction or transcription metric while recording latency, cost, and failure cases.

Curated resources

Reference · AnthropicVisionPractical image input, tiling and token-cost rules.Course · Hugging FaceAudio CourseSpeech recognition, TTS and audio pipelines from scratch.Reference · OpenAIRealtime APIDuplex voice, interruption handling and session state.Course · Hugging FaceComputer Vision CourseOptional depth on vision models and multimodal architectures.Build · IBMDoclingUse it to add messy documents to your existing project.

10

Adaptation: Fine-Tuning & Distillation

The lever you reach for only after prompting, retrieval, and evals have run out of room.

SpecializationEffort L0/17
Done when: Fine-tune one small model and beat the frozen baseline on a held-out set without increasing the critical-failure rate.

Curated resources

Reference · Hugging FacePEFTThe canonical library and method reference for efficient tuning.Read · Microsoft ResearchLoRAThe paper that made fine-tuning affordable.Read · StanfordDirect Preference OptimizationPreference alignment without a separate reward model.Course · DeepLearning.AIFinetuning Large Language ModelsA short, practical first pass at the workflow.Build · UnslothUnslothUse it to test whether a small fine-tune beats your project baseline.

11

Inference, Serving & Hardware

Where token cost and latency are actually decided — essential once you self-host.

SpecializationEffort M0/17
Done when: Serve and load-test one model; report p50 and p95 latency, throughput, memory, cost, and quality change after quantisation.

Curated resources

Reference · vLLMvLLMThe default high-throughput serving engine and its knobs.Read · UC BerkeleyEfficient Memory Management (PagedAttention)Why serving throughput improved by an order of magnitude.Reference · Hugging FaceLLM Inference OptimizationCaching, quantisation and decoding techniques in one place.Reference · OllamaOllamaThe fastest way to get a local model under a load test.Build · ggmlllama.cppQuantise one project candidate and measure the quality you lose.

12

Production, LLMOps & Scaling

Keep reliability, latency, quality, and cost predictable as usage grows.

CoreEffort M0/26
Done when: Deploy with an SLO, cost-per-task budget and alert, rate limit, and rollback path; pass a load test at the declared target concurrency.

Curated resources

Course · Made With MLMLOps CourseDesign, deploy, test and iterate production ML.Read · GoogleRules of Machine LearningBattle-tested priorities for production ML systems.Read · GoogleSite Reliability EngineeringReliability principles that AI systems still need.Course · KubernetesKubernetes BasicsOfficial deployment, scaling and update tutorial.Course · FSDLFull Stack Deep LearningPre-LLM (2022) — take the lifecycle thinking, not the tooling.Build · BerriAILiteLLMAdd a gateway only if your project needs routing, failover, or budgets.Reference · AnthropicBatch ProcessingHalve the bill on anything that does not need an answer now.

13

AI System Design

Turn requirements into architectures with explicit quality, cost, security, and failure trade-offs.

CoreEffort S0/20
Done when: Write and defend a one-page design covering SLOs, trust boundaries, cost, failure modes, rollback, and one rejected alternative.

Curated resources

Course · Stanford CS329SMachine Learning Systems DesignPre-LLM (2022) — still the best framing of production trade-offs.Read · Chip HuyenDesigning Machine Learning SystemsThe modern reference for production ML design.Read · Eugene YanPatterns for LLM SystemsEvals, RAG, caching, guardrails and feedback patterns.Read · Chip HuyenAgentsA rigorous decomposition of agent architectures.Build · Donne MartinSystem Design PrimerUse its prompts to pressure-test your project design.

14

AI Product UX

Make probabilistic systems understandable, controllable, accessible, and pleasant to use.

SpecializationEffort S0/16
Done when: Add streaming, citations, approval, and recovery; complete the critical path by keyboard and screen reader and test it with five users.

Curated resources

Read · Microsoft HAXGuidelines for Human-AI InteractionEvidence-based patterns for usable AI products.Reference · VercelAI SDKStreaming, tool calls and generative UI in practice.Read · W3C WAIIntroduction to Web AccessibilityAccessibility is a product requirement, not polish.Course · ReactLearn ReactOptional — only if the frontend basics are missing.Build · VercelAI Chatbot TemplateReference its streaming and tool UI patterns in your project.

15

Capstone & Staying Current

Prove the skill by shipping one narrow, evaluated, secure, observable system — then keep the loop running.

CoreEffort L0/18
Done when: Deploy to real users; meet the golden-set target, pass red-team and load checks, and publish the demo, design, decisions, and postmortem.

Curated resources

Build · OpenAIOpenAI CookbookSmall, production-oriented implementation recipes.Build · AnthropicAnthropic CookbookTool use, retrieval, classification and evaluation examples.Build · FSDLFull Stack Deep Learning LabsA guided path from model to deployed product.Read · Simon WillisonSimon Willison on LLMsThe most reliable running commentary on what actually changed.Watch · AI EngineerAI EngineerConference talks from teams running this in production.