Volver al CV

Plan de estudios

Study plan: getting into applied AI (2026)

Estás mirando la ruta desde el posicionamiento: Frontend Developer

0/26 recursos con marca y 0/7 gates cumplidos

Perfil frente al plan

Cubiertas

  • TypeScriptAdvanced
  • SQL and databases (SQLite and PostgreSQL)Basic
  • Prompt engineeringIntermediate

Gaps

  1. AI engineering craft (role, trade-offs and working with agents)Esto ya lo tocaste en: United Airlines (Flight Following console), United Airlines (Flight Following console), BBVA Seguros Argentina
  2. Python70.8% of analyzed postings; 78.8% of role postings in Spain
  3. Cloud (AWS first)AWS in 40.3% of postings (the top cloud), Azure 29.6% and GCP 27.4%; 54% of role profiles in Spain declare cloud
  4. Docker1,700 of 6,964 postings; 32.6% of role profiles in Spain declare Docker or Kubernetes
  5. FastAPI680 of 6,964 postings; 14.6% of role profiles in Spain
  6. LLM APIs and structured outputs62.2% of postings mention LLMs; 78.8% of role postings in SpainEsto ya lo tocaste en: United Airlines (Flight Following console), United Airlines (Flight Following console), BBVA Seguros Argentina, Smart Face
  7. Embeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)
  8. RAG (Retrieval-Augmented Generation)39.8% of all postings (53.9% of AI-First ones); 56.7% of role postings in Spain
  9. Vector databases1,647 of 6,964 postings mention vector databases; 8.3% of role profiles in Spain
  10. Agent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in SpainEsto ya lo tocaste en: United Airlines (Flight Following console)
  11. LLM frameworks (LangChain, LangGraph, LlamaIndex)LangChain in 22% of postings; LangChain or LangGraph in 36.3% of role postings in Spain (the real differentiator: 21 times more than the rest of the market)
  12. Model Context Protocol (MCP)14.1% of postings and rising (from 9.9% to 17.6% during 2026); 7.6% of role profiles in SpainEsto ya lo tocaste en: United Airlines (Flight Following console)
  13. LLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baselineEsto ya lo tocaste en: BBVA Seguros Argentina
  14. Observability and monitoring24.5% of AI-First postings; 14.1% of role profiles in Spain mention it in their self-description
  15. Guardrails and agent safety17.9% of AI-First postings ask for guardrails; safety (prompt injection, exfiltration) is tested in interviewsEsto ya lo tocaste en: United Airlines (Flight Following console)
  16. Fine-tuning (specialization)27.3% of AI-First postings mention it but only 3.6% make it a core responsibility: it is not on the critical path

Lo que ya tenés fuera del plan

  • ReactAdvanced
  • Redux ToolkitAdvanced
  • React QueryIntermediate
  • Sass/SCSSIntermediate
  • Responsive designIntermediate
  • Web accessibilityIntermediate
  • AG GridIntermediate
  • ViteIntermediate
  • VitestIntermediate
  • React Testing LibraryIntermediate
  • PlaywrightIntermediate
  • StorybookIntermediate
  • REST APIsAdvanced
  • Git and GitHubAdvanced
  • CI/CD and feature flagsIntermediate
  • Node.jsBasic
  • ExpressBasic
  • AI agentic developmentIntermediate
  • Agile/ScrumIntermediate
  • Code review and mentoringIntermediate

Orden de ataque de roles

  1. AI Product Engineer or AI-native Frontend

    Full-stack frontend is the advantage, not the gap: building the whole LLM product, from API to chat UI. Real must-have: a production LLM product with its own UI. It maps closest to the Application and Business job types, with Forward Deployed as the fastest-growing neighbouring title.

    Fase que lo habilita: Python, minimal backend and cloud

  2. AI Automation Engineer or AI Agent Developer

    The lowest-friction path, since it does not compete with trained ML engineers: agent workflows that automate business processes. Real must-have: an end-to-end agent workflow running and the ability to explain the business process it automates.

    Fase que lo habilita: Python, minimal backend and cloud

  3. Senior SWE on an AI product team

    Joining a team that already ships AI as a senior engineer and learning by osmosis, instead of reselling yourself as an unproven ML junior. Keeps seniority and pay, and it is the most recommended route in the community and the closest to how people actually enter the role.

    Fase que lo habilita: Python, minimal backend and cloud

Fases del plan

Fase 0

Decision and focus

1 week
AI engineering craft (role, trade-offs and working with agents)

Temas

  • Pick the two target roles to aim at
  • Define the repo or repos where the projects will live
  • Decide the horizon (3 intensive months or 6 part-time months)
  • Read the map of the four jobs the AI Engineer title covers and the transition path from frontend

Gate: In writing: the two target roles, the horizon and the name of the portfolio repo.

Estado del gate:

Recursos

Qué es un AI Engineer

Eusebio Graciani (nothiring.me)BlogFreeAbout a 20 minute readEspanol

Article read in full on 2026-10-05; first-party Spain data cross-checked against the international dataset (2026-10-05)

The four jobs the AI Engineer title covers (Model, Platform, Application, Business) and what each demands. Useful for picking target roles. Evaluation crosses all four.

AI engineering craft (role, trade-offs and working with agents)
From Frontend Engineer to AI Engineer (AI Engineering Field Guide)

Alexey Grigorev (DataTalksClub)GitHubFreeAbout a 15 minute readIngles

Read in full on 2026-10-05; the repo is a work in progress and keeps growing (2026-10-05)

The measured path from frontend: backend first (TypeScript or Python), then AI, with the full-stack edge as the differentiator. The TypeScript path is the recommended one; Python, the strongly recommended one.

AI engineering craft (role, trade-offs and working with agents)TypeScript19.3% of postings; the role's second language and a valid path from frontendPython70.8% of analyzed postings; 78.8% of role postings in Spain
Fase 1

Python, minimal backend and cloud

2 to 3 weeks

Practicar el quiz de esta fase en /interview

Habilita roles: AI Product Engineer or AI-native Frontend / AI Automation Engineer or AI Agent Developer / Senior SWE on an AI product team

Python70.8% of analyzed postings; 78.8% of role postings in SpainTypeScript19.3% of postings; the role's second language and a valid path from frontendFastAPI680 of 6,964 postings; 14.6% of role profiles in SpainSQL and databases14.4% of postings; 23.3% of role postings in Spain (lower than in a general tech posting)Docker1,700 of 6,964 postings; 32.6% of role profiles in Spain declare Docker or KubernetesCloud (AWS first)AWS in 40.3% of postings (the top cloud), Azure 29.6% and GCP 27.4%; 54% of role profiles in Spain declare cloud

Temas

  • Modern Python: type hints, dataclasses and typing
  • Virtual environments, Pytest, HTTP and JSON
  • FastAPI to expose endpoints, basic SQL and a data access library
  • Docker to run services
  • At least one cloud (AWS first by demand): compute, storage, secrets and logs
  • TypeScript variant as a first-class alternative: Node, Express or Next.js API routes with the same syllabus

Proyecto: A FastAPI service (or Node equivalent) with two endpoints, behavioural tests, a Dockerfile and a deployment on the chosen cloud. Small on purpose.

Gate: Being able to read and modify someone else's Python backend without translating it to JavaScript in your head, and having the service running, tests green and deployed on the cloud.

Estado del gate:

Recursos

CS50's Introduction to Programming with Python

Harvard CS50 (David Malan)edX y YouTubeFree15h57 in the full YouTube version; 9 weeks of contentIngles (con subtitulos)

Duration and scope verified on 2026-09-28 (2026-09-28)

The best serious Python for someone who already programs: type hints, errors, testing with Pytest and basic SQL.

Python70.8% of analyzed postings; 78.8% of role postings in Spain
Python API Development - Comprehensive Course for Beginners

Sanjeev Thiyagarajan (freeCodeCamp.org)YouTubeFree19h00Ingles

Duration and syllabus verified on 2026-09-28 (2026-09-28)

The best backbone for phase 1: it covers the gate exactly, with FastAPI, SQLAlchemy, Postgres, pytest, Docker and a GitHub Actions CI/CD pipeline.

Python70.8% of analyzed postings; 78.8% of role postings in SpainFastAPI680 of 6,964 postings; 14.6% of role profiles in SpainSQL and databases14.4% of postings; 23.3% of role postings in Spain (lower than in a general tech posting)Docker1,700 of 6,964 postings; 32.6% of role profiles in Spain declare Docker or Kubernetes
FastAPI documentation

Sebastian Ramirez y comunidadDocumentacion oficialFreeSection-by-section readingIngles y Espanol

The official docs, with the advanced tutorial on testing and dependencies. It is the Python framework that shows up most in AI postings.

FastAPI680 of 6,964 postings; 14.6% of role profiles in SpainPython70.8% of analyzed postings; 78.8% of role postings in Spain
Curso COMPLETO de PYTHON desde CERO para PRINCIPIANTES

Brais Moure (MoureDev)YouTubeFree10h07Espanol

Video verified on 2026-09-28 (2026-09-28)

A strong Spanish-language resource for phase 1. The pace assumes zero prior knowledge: useful as a reference or for syntax review.

Python70.8% of analyzed postings; 78.8% of role postings in Spain
AWS Free Tier y documentacion

Amazon Web ServicesDocumentacion oficialFree within free tier limitsPer-service readingIngles

AWS is the most demanded cloud in role postings. For phase 1, free tier compute, storage, secrets and logs are enough.

Cloud (AWS first)AWS in 40.3% of postings (the top cloud), Azure 29.6% and GCP 27.4%; 54% of role profiles in Spain declare cloudDocker1,700 of 6,964 postings; 32.6% of role profiles in Spain declare Docker or Kubernetes
Fase 2

LLM fundamentals and prompt engineering

2 to 4 weeks

Practicar el quiz de esta fase en /interview

Habilita roles: AI Product Engineer or AI-native Frontend / AI Automation Engineer or AI Agent Developer / Senior SWE on an AI product team

LLM APIs and structured outputs62.2% of postings mention LLMs; 78.8% of role postings in SpainPrompt engineering34.7% of postings; 12.9% of role profiles in Spain declare it as promptingEmbeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)

Temas

  • How next-token prediction works, context window, tokens and cost
  • Temperature and top-p, system prompts versus user prompts
  • Structured outputs with JSON schema and few-shot prompting
  • ReAct and Chain of Thought
  • Keys and secrets, limitations and typical failures: hallucinations and prompt injection
  • Light reading on ML fundamentals useful to the craft: embeddings, bias and variance, error analysis and data engineering

Proyecto: A command-line chatbot with conversation memory, structured output over a real LLM API and per-session cost control.

Gate: Being able to explain why a prompt fails and fix it, and to justify the choice of model, temperature and output format for a concrete case.

Estado del gate:

Recursos

ChatGPT Prompt Engineering for Developers

Isa Fulford (OpenAI) y Andrew Ng (DeepLearning.AI)DeepLearning.AIFree to audit; certificate and graded assignment on the PRO plan1h40, 9 lessonsIngles

Course page verified on 2026-09-28 (2026-09-28)

The foundational short course for phase 2: summarizing, inferring, transforming, expanding and memory chatbots, with code examples.

Prompt engineering34.7% of postings; 12.9% of role profiles in Spain declare it as promptingLLM APIs and structured outputs62.2% of postings mention LLMs; 78.8% of role postings in Spain
Let's build GPT: from scratch, in code, spelled out

Andrej KarpathyYouTubeFree1h56Ingles

Duration verified on 2026-09-28 (2026-09-28)

The most honest and cheap explanation of why an LLM works: next-token prediction and attention built in small code.

LLM APIs and structured outputs62.2% of postings mention LLMs; 78.8% of role postings in SpainEmbeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)
Prompt Engineering Guide

DAIR.AI (Elvis Saravia y colaboradores)GitHubFree, MIT licenseModular reference guideIngles (con traducciones)

78,697 stars and last push 2026-03-11, verified on 2026-09-28 (2026-09-28)

The permanent reference: ReAct, Chain of Thought, prompt security and use-case guides.

Prompt engineering34.7% of postings; 12.9% of role profiles in Spain declare it as prompting
Claude Academy: AI Fluency y AI capabilities and limitations

Anthropic (equipo de educacion)Anthropic (Claude Academy)FreeAI Fluency: 4h and 14 lessons; AI capabilities and limitations: 3h30 and 13 lessonsIngles

Platform and durations verified on 2026-09-28 (2026-09-28)

Covers what an LLM can and cannot do, limitations and human-agent collaboration. Usage-side, not building-side.

LLM APIs and structured outputs62.2% of postings mention LLMs; 78.8% of role postings in SpainPrompt engineering34.7% of postings; 12.9% of role profiles in Spain declare it as prompting
Fase 3

Embeddings, vector DBs and RAG

4 to 6 weeks

Practicar el quiz de esta fase en /interview

Habilita roles: AI Product Engineer or AI-native Frontend / AI Automation Engineer or AI Agent Developer / Senior SWE on an AI product team

RAG (Retrieval-Augmented Generation)39.8% of all postings (53.9% of AI-First ones); 56.7% of role postings in SpainVector databases1,647 of 6,964 postings mention vector databases; 8.3% of role profiles in SpainEmbeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)

Temas

  • Embeddings and semantic search
  • Vector databases: Chroma to learn, then pgvector, Qdrant or Weaviate
  • Chunking and overlap, hybrid retrieval and re-ranking
  • Grounding answers and retrieval metrics
  • Turning text, PDF, HTML and images into something the model can read, and keeping indexing pipelines clean and current
  • Choosing between a vector index, a knowledge graph or a semantic layer over structured data
  • Typical failures: bad chunking, irrelevant retrieval, unsupported answers

Proyecto: RAG over your own documents with an evaluation suite and error analysis: at least 100 real queries read, failures labelled by hand and a published taxonomy with a before and after number for each change.

Gate: The system answers with cited sources, the evaluation suite runs in CI, the failure taxonomy is published with its numbers, and every chunking and retrieval decision can be explained.

Estado del gate:

Recursos

Retrieval Augmented Generation (RAG)

Zain Hasan (Together.ai y University of Toronto) con DeepLearning.AIDeepLearning.AI (tambien en Coursera)Free to audit; certificate and graded assignments on the PRO plan26h03, 5 modules and 49 lessonsIngles

Exact duration and pricing model verified on 2026-09-28 (2026-09-28)

The most complete resource for phase 3 and the one that matches job postings best: retrievers, BM25, hybrid search, chunking, evaluation and production.

RAG (Retrieval-Augmented Generation)39.8% of all postings (53.9% of AI-First ones); 56.7% of role postings in SpainVector databases1,647 of 6,964 postings mention vector databases; 8.3% of role profiles in SpainLLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Production RAG with LangChain & Vector Databases (Full Course)

freeCodeCamp.orgYouTubeFree7h38Ingles

Published 2026-05-26 and verified on 2026-09-28 (2026-09-28)

The fully free, code-first alternative, ideal for the phase's flagship project: a complete production RAG pipeline.

RAG (Retrieval-Augmented Generation)39.8% of all postings (53.9% of AI-First ones); 56.7% of role postings in SpainLLM frameworks (LangChain, LangGraph, LlamaIndex)LangChain in 22% of postings; LangChain or LangGraph in 36.3% of role postings in Spain (the real differentiator: 21 times more than the rest of the market)Vector databases1,647 of 6,964 postings mention vector databases; 8.3% of role profiles in Spain
RAG: Adapta la IA a tu información

Beto Quiroga (EDteam)EDteam15 USD one-time purchase or premium subscription; module 1 is free2h06, 3 modulesEspanol

4.6 rating with 111 reviews and price verified on 2026-09-28; 5 reviews read in full (2026-09-28)

The most solid Spanish-language resource in the research: cheap, with real reviews read, focused on vectors, semantic search and common pitfalls.

RAG (Retrieval-Augmented Generation)39.8% of all postings (53.9% of AI-First ones); 56.7% of role postings in SpainEmbeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)
Hugging Face NLP Course

Hugging Face (Abubakar Abid, Ben Burtenshaw, Matthew Carrigan y otros)Hugging Face LearnFree and ad-free12 chapters; for RAG take chapters 1 to 4 and 10 to 12Ingles (traduccion al espanol en progreso)

Course page verified on 2026-09-28 (2026-09-28)

The embeddings and transformer mechanics phase 3 needs. Longer than necessary for RAG alone: trim by chapter.

Embeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)
Fase 4

Agents, tool calling and MCP

4 to 6 weeks

Practicar el quiz de esta fase en /interview

Habilita roles: AI Product Engineer or AI-native Frontend / AI Automation Engineer or AI Agent Developer / Senior SWE on an AI product team

Agent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in SpainModel Context Protocol (MCP)14.1% of postings and rising (from 9.9% to 17.6% during 2026); 7.6% of role profiles in SpainLLM frameworks (LangChain, LangGraph, LlamaIndex)LangChain in 22% of postings; LangChain or LangGraph in 36.3% of role postings in Spain (the real differentiator: 21 times more than the rest of the market)AI engineering craft (role, trade-offs and working with agents)

Temas

  • Function calling and tool design
  • Agentic patterns: ReAct, Plan-and-Execute, Reflection
  • Short and long term memory, error handling, timeouts and retry with backoff
  • Cost control and structured outputs
  • Model Context Protocol and building your own MCP server
  • How much autonomy to give the agent and when to step in: the craft of programming with agents

Proyecto: An agent with real tools and error handling, ideally with your own published MCP server. This is where prior experience using MCPs and skills as a user becomes interview material.

Gate: The agent completes a real multi-step flow, with error handling and measured cost, and the tool design choices can be explained.

Estado del gate:

Recursos

Hugging Face Agents Course

Hugging Face (Ben Burtenshaw, Sergio Paniego, Thomas Simonini, Pedro Cuenca y otros)Hugging Face LearnFree; the certification process is also free, with a final challenge scored against GAIA4 units plus 3 bonus ones; 3 to 4h per weekIngles (con traducciones)

32,971 stars on its repo and free certification verified on 2026-09-28 (2026-09-28)

The backbone of phase 4 and the only free certificate that requires proof: agents, tools, smolagents, LangGraph, Agentic RAG and the GAIA benchmark.

Agent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in SpainModel Context Protocol (MCP)14.1% of postings and rising (from 9.9% to 17.6% during 2026); 7.6% of role profiles in SpainLLM frameworks (LangChain, LangGraph, LlamaIndex)LangChain in 22% of postings; LangChain or LangGraph in 36.3% of role postings in Spain (the real differentiator: 21 times more than the rest of the market)
MCP: Build Rich-Context AI Apps with Anthropic

Elie Schoppik (Anthropic)DeepLearning.AIFree during the platform beta; certificate and assignment on PRO1h58, 11 lessonsIngles

Course page verified on 2026-09-28 (2026-09-28)

The direct resource for the gate's own MCP server: client-server architecture, FastMCP, MCP Inspector and remote servers.

Model Context Protocol (MCP)14.1% of postings and rising (from 9.9% to 17.6% during 2026); 7.6% of role profiles in SpainAgent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in Spain
AI Agents in LangGraph

Harrison Chase (LangChain) y Rotem Weiss (Tavily)DeepLearning.AIFree to audit; certificate on PRO1h32, 9 lessonsIngles

Course page verified on 2026-09-28 (2026-09-28)

Controlled agent design, taught by the framework's creator: persistence, streaming and human-in-the-loop.

Agent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in SpainLLM frameworks (LangChain, LangGraph, LlamaIndex)LangChain in 22% of postings; LangChain or LangGraph in 36.3% of role postings in Spain (the real differentiator: 21 times more than the rest of the market)
Agentic AI

Andrew Ng (DeepLearning.AI)DeepLearning.AIVideos are free to audit; certificate, assignments and labs on PRO9h55, 31 lessonsIngles

Page and PRO plan FAQ verified on 2026-09-28 (2026-09-28)

The course that ties phase 4 to phase 5 best: reflection, tool use, planning, multi-agent, MCP, evals and error analysis.

Agent patterns and tool calling55.4% of postings (71.6% of AI-First ones); 73.3% of role postings in SpainLLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Fase 5

Evals, observability and production

3 to 4 weeks

Practicar el quiz de esta fase en /interview

Habilita roles: AI Product Engineer or AI-native Frontend / AI Automation Engineer or AI Agent Developer / Senior SWE on an AI product team

LLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baselineObservability and monitoring24.5% of AI-First postings; 14.1% of role profiles in Spain mention it in their self-descriptionGuardrails and agent safety17.9% of AI-First postings ask for guardrails; safety (prompt injection, exfiltration) is tested in interviews

Temas

  • Evaluation datasets, automatic graders and human-in-the-loop
  • Regression detection across prompt or model versions and prompt versioning
  • Tracing and logging, rate limiting and caching
  • Cost and latency monitoring and usage dashboards
  • Guardrails, responding to prompt injection and model failures, drift detection
  • Statistical regression testing calibrated to the risk of the error
  • Cost and latency reduction: model choice, distillation, simplifying the workflow

Proyecto: An assistant or chatbot with persistent memory, deployed with its own UI, evals in the release pipeline and a second published round of error analysis. It is the project where a frontend profile hits hardest.

Gate: The system is deployed with a live URL, evals run in CI and there is at least one detected and documented regression case.

Estado del gate:

Recursos

Your AI Product Needs Evals y AI Evals Email Course

Hamel Husain (ex Airbnb y GitHub) con Shreya ShankarBlog personal y curso por emailFree (the email course includes 2 ebooks)The post: a 1 to 2h read; the email course: 17 lessonsIngles

Free material verified on 2026-09-28 (2026-09-28)

The best resource in the research for signal-to-cost ratio. This is where error analysis lives: reading traces, hand-labelling failures and publishing the taxonomy with numbers.

LLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Quality and Safety for LLM Applications

Bernease Herman (WhyLabs)DeepLearning.AIFree to audit; certificate on PRO2h16, 7 lessonsIngles

Course page verified on 2026-09-28 (2026-09-28)

Hallucinations (SelfCheckGPT), data leakage, prompt injection and passive and active monitoring: the quality and safety side of phase 5.

Guardrails and agent safety17.9% of AI-First postings ask for guardrails; safety (prompt injection, exfiltration) is tested in interviewsLLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Langfuse (documentacion y plan Hobby)

LangfusePlataforma open sourceFree Hobby plan (50,000 units per month); free self-hostingDocumentation and guides to readIngles

Pricing and tiers verified on 2026-09-28 (2026-09-28)

Agent traces, token cost, prompt versioning, evaluation datasets and LLM-as-judge, with a real free tier for learning and for the portfolio.

Observability and monitoring24.5% of AI-First postings; 14.1% of role profiles in Spain mention it in their self-descriptionLLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Bonus Unit 2: Agent Observability and Evaluation

Hugging Face Agents CourseHugging Face LearnFreeA bonus unit of a few hoursIngles

Bonus unit content verified on 2026-09-28 (2026-09-28)

The shortcut if you are already taking the agents course: it covers phase 5 evals without switching platforms.

Observability and monitoring24.5% of AI-First postings; 14.1% of role profiles in Spain mention it in their self-descriptionLLM system evaluation and error analysis59.7% of AI-First postings ask for evaluation; 51.5% to 68.9% of role postings in Spain mention it. It is the differentiator: RAG and agents are the baseline
Fase 6

Optional specialization

After the phase 3 and 5 gates
Fine-tuning (specialization)27.3% of AI-First postings mention it but only 3.6% make it a core responsibility: it is not on the critical pathGuardrails and agent safety17.9% of AI-First postings ask for guardrails; safety (prompt injection, exfiltration) is tested in interviews

Temas

  • Fine-tuning and LoRA for model depth
  • Multimodal
  • Advanced RAG with knowledge graphs
  • AI safety and red teaming
  • Vertical domain: legal, health, finance

Gate: Only with the phase 3 and 5 gates met. Fine-tuning and classical ML stay off the critical path.

Estado del gate:

Recursos

LLM Fine-Tuning Course: From Supervised FT to RLHF, LoRA, and Multimodal

freeCodeCamp.orgYouTubeFree11h56Ingles

Published 2026-03-10 and verified on 2026-09-28 (2026-09-28)

The most direct option for the fine-tuning specialization: supervised, RLHF, LoRA and multimodal. Only after the phase 3 and 5 gates.

Fine-tuning (specialization)27.3% of AI-First postings mention it but only 3.6% make it a core responsibility: it is not on the critical path
Practical Deep Learning for Coders

Jeremy Howard (fast.ai)fast.aiFree, no account or card neededPart 1: 9 lessons of about 90 minutes eachIngles

Page verified on 2026-09-28 (2026-09-28)

Only if the specialization needs model depth. For applied AI engineering it is truly optional: no certificate, top-down approach.

Fine-tuning (specialization)27.3% of AI-First postings mention it but only 3.6% make it a core responsibility: it is not on the critical pathEmbeddings and semantic search22.9% of AI-First postings ask for embeddings (the most requested ML-adjacent skill after fine-tuning)
Red Teaming LLM Applications

Matteo Dora y Luca Martial (Giskard)DeepLearning.AIFree to audit; certificate on PRO1h29, 7 lessonsIngles

Course page verified on 2026-09-28 (2026-09-28)

The AI safety and red teaming option on the specialization menu: vulnerabilities, prompt injection and manual and automated red teaming.

Guardrails and agent safety17.9% of AI-First postings ask for guardrails; safety (prompt injection, exfiltration) is tested in interviews

Qué no hacer

  • Do not pay for an expensive ML/AI bootcamp thinking it is the entry door.
  • Do not start with linear algebra, calculus, PyTorch or training networks if the goal is applied AI engineering: that path trains researchers.
  • Do not complete the whole Deep Learning Specialization before touching RAG.
  • Do not present yourself as a Prompt Engineer: the title got absorbed into other roles.
  • Do not apply to junior ML Engineer or junior MLOps roles with this plan: they are different profiles and need more than the plan covers.
  • Do not add skills to the CV without real, defensible use. It applies to AI the same way it applies to frontend.

Señales de salida

  • Phase 3 completed with the RAG project deployed: you can start applying to automation and agent roles and to positions asking for RAG with a UI.
  • Phase 5 completed with the assistant deployed and evals in CI: coverage of all three target roles, including AI Product Engineer.
  • Every project goes into the job search with its error analysis published: a failure taxonomy with before and after numbers. It is the answer to the evaluation question in interviews.
  • Projects only enter the CV once deployed with a live URL.
  • Shifting the CV positioning toward AI is a separate decision. If taken, also search for AI Product Engineer, Full Stack AI Engineer and Forward Deployed: searching only for AI Engineer misses most of the talent and most postings.

Material de investigación

Lecturas del dominio, ordenadas para profundizar. No forman parte del contrato del plan ni del perfil.

  1. Qué es un AI Engineer (nothiring.me)

    The four jobs the AI Engineer title covers (Model, Platform, Application, Business) and what each demands, crossing Spain data with Grigorev's international dataset.

  2. Quién es el AI Engineer en España (nothiring.me)

    Who holds the role, where they come from, how much experience they bring and why the entry door barely exists. The real route from software.

  3. Cuánto cobra un AI Engineer en España (nothiring.me)

    Pay bands for the role, the product versus consultancy gap for the same title, and how much competition each posting gathers.

  4. AI Engineering Field Guide (Alexey Grigorev)

    Repo with 6,964 job descriptions analysed, interview patterns from 100+ sources, learning paths by professional background and a portfolio guide.

  5. Mapa de habilidades de ingeniería de IA (Andrew Ng)

    The role's four top-level skills and the six pieces of building and deploying AI applications, drawn from 10,000+ job postings.

  6. Your AI Product Needs Evals (Hamel Husain)

    The error analysis method: read real traces, label failures by hand and publish the taxonomy with numbers. It is what job postings mean by evaluation.

  7. The AI Engineering Stack (Chip Huyen)

    The definition that separates AI engineering from ML: adapt and evaluate models, not build them. Prompting versus fine-tuning as two paths.