3-month plan: the thorough path
Twelve weeks through all sixteen topics on one project, an assistant over documents you know, that grows from a single LLM call into an evaluated, guarded and deployed RAG agent with a small fine-tuned router. About 8 hours a week. The topics follow the learning path; other plans: 1-month plan · 2-month plan · 6-month plan.
An unofficial study plan built from the public posts of Bashiri Smith and Melvin Vivas, not affiliated with or endorsed by them. Items under Fill the gap are not from the creators: they cover skills their posts don't. Track your progress in the library app; ticks are saved in your browser.
Week 1: Map the role and plan
Goal Explain what AI engineers build and get hired for, rate your skills, and write a capstone brief with 20 real test questions. (about 7.5 h in total)
Core
- The Complete AI Engineer Roadmap for 2026 (Exact Courses + Step-by-Step) · Video · youtube.com · the creator's own
Do: Whole video (30 min). Note the five skill areas (LLMs & RAG, agents, ops & evaluation, safety & ethics, ML foundations) and the course he names for each; this plan covers all five but puts a short ML-foundations week first, where he leaves it for last · about 45 min. - AI Engineering Field Guide (Alexey Grigorev) · Repo · github.com
Do: Read 'The AI Engineer Role' (skills analysis, responsibilities, use cases, reality vs. job postings) and 'Learning Paths: From Backend Engineer' (or From Frontend Engineer). Bookmark 'Interview Preparation' for week 12 · about 2 h.
From the creators' posts
- 3-Month Roadmap to AI Engineering for Experienced Software Engineers · Bashiri Smith, Facebook · 0:07
- The 3 Levels of AI Engineering: LLM Apps → Production → Agentic Systems · Bashiri Smith, Facebook · 1:13
- How to Relearn LLMs & RAG in 2026: A 7-Step Roadmap with Free Resources · Bashiri Smith, Facebook · 1:54
- 6 Steps to Build a Resume Project That Lands an AI Engineering Job · Bashiri Smith, Facebook · 2:47
Build
- Set up your toolkit (Python 3.11+, Git, a hosted LLM API key, Ollama or LM Studio, Google Colab) and write a one-page capstone brief: an assistant over documents you know well, ideally solving a real problem at a company you'd like to join, who would use it, and 20 real questions with the correct answer and source page (your first eval set). Rate yourself 1-5 on each skill area. (about 4.5 h) Idea from 3-Month Roadmap to AI Engineering for Experienced Software Engineers.
Optional
- AI Engineering Skills Map: Building and Deploying AI Applications · Article · x.com
- 5 books to upgrade from software engineer into Ai/ML engineer · PDF · docs.google.com · the creator's own
Week 2: How neural networks learn
Goal Explain how a neural network learns (loss, gradient descent, backpropagation) and train a tiny network you wrote from scratch. (about 8.5 h in total)
Core
- 3Blue1Brown Neural Networks series · Video · youtube.com
Do: Chapters 1-4 only (about 62 min: what a neural network is, gradient descent, backpropagation intuition and calculus). Chapters 5-7 on transformers are a good extra before week 3. - Neural Networks: Zero to Hero (Andrej Karpathy) · Course · karpathy.ai
Do: Lecture 1 only: 'The spelled-out intro to neural networks and backpropagation: building micrograd' (2 h 26 min). Code along in a notebook instead of just watching; the other lectures are optional later · about 4 h.
From the creators' posts
- Software Engineering Fundamentals Before AI Engineering · Bashiri Smith, Facebook · 0:49
- AI Engineer Mistakes Pt. 3: Don't Grind Math First, Choose Embeddings Carefully · Bashiri Smith, Facebook · 0:31
- Run Local Models on a Free GPU with Google Colab (T4) · Melvin Vivas, X · 0:54
Build
- Finish your micrograd notebook on your own in Google Colab: train a small MLP on a toy dataset, plot the loss going down, and add a short note to your capstone repo explaining gradient descent and backpropagation in your own words. (about 1.5 h)
Fill the gap (not from the creators)
- Pydantic docs: Models · Docs · pydantic.dev
Type-hinted models and validation underpin FastAPI, structured-output SDKs (OpenAI parse, Instructor) and agent frameworks, so this is the Python skill AI work leans on most. (about 1 h) - LLM Course: LLM Fundamentals (Maxime Labonne) · Course · github.com
The course's refresher on the maths, Python and neural networks behind LLMs. Its author suggests using it as a reference when needed rather than starting there. (about 15 min)
Optional
- Mathematics for Machine Learning · Book · mml-book.github.io
Optional, to fill the gap (not from the creators)
- Stanford CME 295, Lecture 1: Transformers · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 1 of 9): NLP background, tokenization, embeddings, word2vec, RNNs and LSTMs, then attention and the Transformer. (about 1.5 h)
Week 3: LLM fundamentals and APIs
Goal Explain tokens, training and context windows, call hosted and local LLMs from Python, and choose a model on your own questions. (about 9.5 h in total)
Core
- [1hr Talk] Intro to Large Language Models (Andrej Karpathy) · Video · youtube.com
Do: Whole talk (1 h) · about 1.5 h. - OpenAI, Claude & Gemini API Tutorial in Python (Machine Learning Plus) · Article · machinelearningplus.com
Do: Whole tutorial with its three exercises (first call, error handling, a unified wrapper, multi-turn, streaming, cost estimate). The model names are dated; use current ones · about 1.5 h. - Artificial Analysis · Website · artificialanalysis.ai
Do: Use the model comparison (intelligence, speed, price) to shortlist 2-3 hosted models for your capstone, then check them on your own questions · about 30 min.
From the creators' posts
- How GPT Works: From Token Embeddings to Multi-Head Attention · Melvin Vivas, X · 15:11
- Running LLMs Locally Without an Expensive Rig · Melvin Vivas, X
- Test models on your own workflow, not public benchmarks · Melvin Vivas, X
Build
- Capstone v0, a long-context baseline: a small Python CLI or FastAPI endpoint ask(question) that puts one or two of your documents straight into the prompt and calls a hosted model and a local model (Ollama or LM Studio). Run your 20 questions through 2-3 models, log tokens, latency and estimated cost per call, and record which answers are right. (about 4 h) Idea from Test models on your own workflow, not public benchmarks.
Fill the gap (not from the creators)
- Reasoning models guide (OpenAI API docs) · Docs · developers.openai.com
Explains that reasoning tokens are billed as output and take up context, the effort levels from none to max and their speed/cost trade-off, cost control and prompting advice; the same ideas apply to Claude and Gemini thinking budgets. (about 30 min) - LLM Course: LLM Architecture (Maxime Labonne) · Course · github.com
How modern LLMs turn text into tokens, pass them through attention layers and generate text with different sampling strategies, with a short list of the best explainers. (about 15 min) - LLM Course: Pre-training (Maxime Labonne) · Course · github.com
How base models are pre-trained: data preparation, distributed training and scaling. Background, not a skill you need to use: it explains where the models you call or fine-tune come from. (about 15 min) - LLM Course: Running LLMs (Maxime Labonne) · Course · github.com
Using LLMs through APIs or running open models locally, plus prompt engineering and structuring outputs, with links to the tools for each. (about 15 min) - Phase 10: LLMs from Scratch (AI Engineering from Scratch, Rohit Ghumare) · Course · github.com
24 build-it-yourself lessons on how LLMs work, from tokenizers through pre-training and scaling. Skim the lesson list and bookmark the ones behind this week's topics. (about 15 min)
Optional
- OpenAI with Python: A Step-by-Step Guide for Beginners (George Shakan) · Video · youtube.com
Optional, to fill the gap (not from the creators)
- Stanford CME 295, Lecture 2: Transformer-Based Models and Tricks · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 2 of 9): attention variants (MHA, MQA, GQA), position embeddings and RoPE, and Transformer-based models such as BERT and its derivatives. (about 2 h) - Stanford CME 295, Lecture 3: Large Language Models · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 3 of 9): what an LLM is, mixture of experts, context length, temperature, sampling, prompting, in-context learning, chain of thought and self-consistency. (about 2 h)
Week 4: Prompts and coding agents
Goal Write testable prompts that return validated JSON, and use a coding agent with an AGENTS.md and plan mode. (about 9 h in total)
Core
- ChatGPT Prompt Engineering for Developers · Course · deeplearning.ai
Do: Whole course (1 h 40 min, 9 lessons); run the notebooks with the model you picked in week 3 · about 2 h. - Prompting best practices (Claude Platform Docs) · Docs · platform.claude.com
Do: The 'Techniques for all current models' part: general principles, output and formatting, tool use and agentic systems. Skip the model-specific and migration sections · about 45 min. - AGENTS.md · Docs · agents.md
Do: The format and examples (10 min), then write one for your repo · about 15 min.
From the creators' posts
- Why Plan Mode Still Matters in AI Coding Agents (Codex, Claude, Cursor) · Melvin Vivas, X · 5:54
- Building and Deploying a Full App with the Codex App, MCP Servers and Skills · Melvin Vivas, X · 3:53
- Cursor 3.3: See Which Rules, Skills, MCPs and Subagents Use Your Context · Melvin Vivas, X · 0:25
- Use HTML for LLM Outputs, Markdown for Context · Melvin Vivas, X
Build
- Capstone v0.5: make ask() return {answer, sources, confidence} as JSON validated with Pydantic, with one retry when the output doesn't parse, and keep prompts in versioned files. Add an AGENTS.md that describes your stack and how to run the tests, then use a coding agent in plan mode to design and implement one feature (for example a /ask endpoint with tests) and review the diff before you merge it. (about 4 h)
Fill the gap (not from the creators)
- Function calling guide (OpenAI API docs) · Docs · developers.openai.com
The reference for defining tools with JSON Schema, strict mode, tool_choice, parallel calls and the call-execute-return loop; OpenRouter, llama.cpp and LM Studio (used throughout the library) serve the same OpenAI-compatible format. (about 1 h) - Effective context engineering for AI agents (Anthropic Engineering) · Article · anthropic.com
Covers the core moves: minimal high-signal system prompts, just-in-time retrieval, compaction, structured note-taking and sub-agents. (about 30 min)
Optional
- Building Systems with the ChatGPT API · Course · deeplearning.ai
- Prompt Engineering Full Course (Tech With Tim) · Video · youtube.com
- Parallel Coding Agents with Git Worktrees and Playwright MCP · Melvin Vivas, X
Optional, to fill the gap (not from the creators)
- Context Engineering (AI Engineering from Scratch, Rohit Ghumare) · Tutorial · github.com
A hands-on lesson: build a context assembler that budgets the window between instructions, retrieved documents, tools and conversation memory. (about 1.5 h) - Stanford CME 295, Lecture 6: LLM Reasoning · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 6 of 9): reasoning models, reinforcement learning for reasoning, GRPO and scaling. (about 2 h)
Week 5: Embeddings and vector search
Goal Embed, store and search text in a vector database, explain HNSW, and pick an embedding model by measuring retrieval on your data. (about 8 h in total)
Core
- Word Embedding and Word2Vec, Clearly Explained!!! (StatQuest with Josh Starmer) · Video · youtube.com
Do: Whole video (16 min) for the intuition of meaning as vectors · about 30 min. - Vector Databases: from Embeddings to Applications (DeepLearning.AI) · Course · deeplearning.ai
Do: Whole course (1 h 5 min, 8 lessons): embeddings, distance metrics, approximate nearest neighbours/HNSW, sparse, dense and hybrid search · about 1.5 h. - Sentence Transformers · Tool · sbert.net
Do: Quickstart, then the 'Semantic Search' and 'Retrieve & Re-Rank' guides (about 1 h).
From the creators' posts
- How RAG Finds the Right Document Fast: Graph-Based Vector Search · Bashiri Smith, Facebook · 1:08
- Fine-tuning all-MiniLM-L6-v2 to Beat OpenAI Embeddings · Melvin Vivas, X
- RAG vs CAG: Retrieval vs Cache Augmented Generation Explained · Bashiri Smith, Facebook · 1:31
Build
- Capstone v1, retrieval: split your documents into chunks, embed them with all-MiniLM-L6-v2 (Sentence Transformers) and with one API embedding model, store the vectors in pgvector or Chroma, and write search(query, k). For each of your 20 questions check whether the right source is in the top 5 (hit@5) and keep the better model. (about 4 h) Idea from Fine-tuning all-MiniLM-L6-v2 to Beat OpenAI Embeddings.
Fill the gap (not from the creators)
- Step-by-Step Guide to Choosing the Best Embedding Model (Weaviate) · Article · weaviate.io
A 4-step process: define the use case, shortlist from MTEB by task, size, dimensions and max tokens, then build a small hand-labelled set (50-100 documents plus test queries) from your own data and compare models on precision and recall. (about 30 min)
Optional
- What is a Vector Database? Powering Semantic Search & AI Applications (IBM Technology) · Video · youtube.com
- pgvector · Tool · github.com
Week 6: RAG that holds up
Goal Build RAG that answers only from your documents with citations, and improve it with chunking, hybrid search and reranking. (about 8.5 h in total)
Core
- LangChain: RAG From Scratch (YouTube playlist) · Video · youtube.com
Do: Parts 1-9 (about 52 min: indexing, retrieval, generation and query translation: multi-query, RAG-Fusion, decomposition, step-back, HyDE), coding along with the notebooks in the rag-from-scratch repo. - 5 Levels of Text Splitting for Retrieval (Greg Kamradt) · Video · youtube.com
Do: Levels 1-3 (character, recursive and document-specific splitting); skim levels 4-5 · about 45 min. - Build Production-Ready Retrieval RAG Pipeline in LangChain | Hybrid Search (BM25), Re-ranking & HyDE (Venelin Valkov) · Video · youtube.com
Do: Whole video (15 min): BM25 hybrid search, re-ranking and HyDE · about 30 min.
From the creators' posts
- How RAG Works Under the Hood: Chunking, Embedding, Storage, Retrieval · Bashiri Smith, Facebook · 1:12
- RAG Chunking: Balancing Chunk Size for Precision and Context · Bashiri Smith, Facebook · 0:29
- Fixing RAG Ranking Problems with a Cross-Encoder Re-ranker · Bashiri Smith, Facebook · 1:10
- Why Blind Top-K Retrieval Hurts RAG, and What to Use Instead · Bashiri Smith, Facebook · 1:11
Build
- Capstone v1, RAG: send the top chunks to the model with instructions to answer only from them, cite the source and say "I don't know" otherwise. Then upgrade retrieval with BM25 keyword search (hybrid), a wider candidate set reranked by a cross-encoder and a relevance threshold, and record hit@5 and answer correctness on your 20 questions after each change. (about 4 h) Idea from 3 Resume-Ready RAG Projects: Hybrid Search, Multi-Modal Docs and Agentic RAG.
Fill the gap (not from the creators)
- Contextual Retrieval (Anthropic Engineering) · Article · anthropic.com
Shows how adding LLM-generated context to each chunk before embedding and BM25 indexing, plus reranking, cut failed retrievals by up to 67%, with a runnable cookbook. (about 30 min) - LLM Course: RAG (Maxime Labonne) · Course · github.com
Orchestrators, retrievers, memory and evaluation for retrieval-augmented generation; the sections just before and after cover vector storage and advanced RAG. (about 15 min) - Phase 11: LLM Engineering (AI Engineering from Scratch, Rohit Ghumare) · Course · github.com
17 lessons on putting LLMs to work: prompt and context engineering, RAG, fine-tuning and evals. Skim the lesson list; the RAG lessons fit this week. (about 15 min)
Optional
- How to become an expert in RAG (BASWE AI Engineer Field Guide) · PDF · drive.google.com · the creator's own
- Advanced Retrieval for AI with Chroma (DeepLearning.AI) · Course · deeplearning.ai
- 3 Reasons a "Correct" RAG Pipeline Still Fails in Production · Bashiri Smith, Facebook · 1:03
Optional, to fill the gap (not from the creators)
- Advanced RAG (AI Engineering from Scratch, Rohit Ghumare) · Tutorial · github.com
A hands-on lesson: hybrid search that combines BM25 keyword matching with vector search, plus a cross-encoder reranker and better chunking. (about 1.5 h) - Production RAG with LangChain & Vector Databases (freeCodeCamp) · Course · youtube.com
Whole course: RAG from basics to production: debugging, hybrid search, observability, optimization, scaling, hosting and security. (about 7.5 h)
Week 7: Agents, tools and MCP
Goal Build a bounded tool-using agent, serve a tool from your own MCP server, and pick the simplest agent pattern that works. (about 9 h in total)
Core
- BASWE: Agent Cheatsheet for AI Engineers · PDF · drive.google.com · the creator's own
Do: The agentic design patterns PDF: Parts 1-2 (do the tool-use exercise) and sections 4.1-4.6 · about 2 h. - Model Context Protocol (MCP) · Docs · modelcontextprotocol.io
Do: 'What is MCP' and 'Architecture', then the 'Build an MCP server' Python quickstart; test your server with the MCP Inspector or your agent · about 1.5 h.
From the creators' posts
- 5 Agentic Design Patterns: ReAct, Planning, Reflection, Routing, Multi-Agent · Bashiri Smith, Facebook · 0:11
- LangChain vs. LangGraph Explained with One RAG Chatbot · Bashiri Smith, Facebook · 1:13
- Running Local Models for Agents: Tool Use, Context and Quantization · Melvin Vivas, X · 1:45
- Agent Harness Efficiency: Scaffolding Beats MCP vs. CLI · Melvin Vivas, X
Build
- Capstone v2, a bounded agent: wrap your retriever as a search_docs tool served from your own MCP server, add one action tool (for example create_ticket, writing to a file or a GitHub issue) that runs only after you approve it, let the agent rewrite the query and search again when results are weak, cap it at 5 steps, and log every tool call. (about 3.5 h) Idea from LangChain vs. LangGraph Explained with One RAG Chatbot.
Fill the gap (not from the creators)
- Building effective agents (Anthropic Engineering) · Article · anthropic.com
The canonical guide to workflows vs agents (prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer) and to starting with the simplest solution, with linked cookbook code. (about 1 h) - Phase 13: Tools and Protocols (AI Engineering from Scratch, Rohit Ghumare) · Course · github.com
From function calling and tool schemas to MCP and other agent protocols, with security and governance. The phase page suggests shorter routes through its 45 lessons. (about 15 min) - Phase 14: Agent Engineering (AI Engineering from Scratch, Rohit Ghumare) · Course · github.com
Agents built from first principles, starting with the agent loop. The phase page lists two shorter routes if you do not need all of it. (about 15 min)
Optional
- What Are AI Agents? (IBM Technology) · Video · youtube.com
- Master All 20 Agentic Design Patterns (Mark Kashef) · Video · youtube.com
- Hugging Face Agents Course · Course · huggingface.co
Optional, to fill the gap (not from the creators)
- Build an MCP Server (AI Engineering from Scratch, Rohit Ghumare) · Tutorial · github.com
A hands-on lesson: the same stateless notes server built in Python and in TypeScript, using only the standard library for the protocol core. (about 1.5 h) - Stanford CME 295, Lecture 7: Agentic LLMs · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 7 of 9): retrieval-augmented generation, advanced RAG techniques, function calling, agents and the ReAct framework. (about 2 h) - Agentic AI: Complete Course for Beginners (freeCodeCamp) · Course · youtube.com
Optional and long (24 hours): a complete beginner course on agentic AI, for after the plan. (about 24.5 h) - LangGraph Complete Course for Beginners (freeCodeCamp) · Course · youtube.com
Whole course: LangGraph state graphs built up over five practice graphs and four AI agents, ending with a RAG agent. (about 3 h) - Intro to MCP Servers: Model Context Protocol with Python (freeCodeCamp) · Course · youtube.com
Whole course: building and connecting MCP servers in Python. (about 1 h)
Week 8: Evals and guardrails
Goal Measure retrieval and answers with a golden set and a checked LLM judge, and defend against prompt injection, unsafe input and PII leaks. (about 9.5 h in total)
Core
- Evaluation Field Guide (baswe.ai engineer accelerator – Ops and Evaluation module) · PDF · drive.google.com · the creator's own
Do: Eval fundamentals, metrics, LLM-as-a-judge, RAG evaluation and agent trajectory evaluation (about half of the 41 pages); keep the production-evals chapter for week 9. - promptfoo · Repo · github.com
Do: Getting started (config, assertions, model-graded metrics), then the red-team quickstart for prompt injection and jailbreaks · about 1.5 h. - GliGuard Moderation Colab Notebook · Tool · colab.research.google.com · the creator's own
Do: Run it once in Colab and note how prompts and responses get labelled · about 30 min.
From the creators' posts
- How to Evaluate a RAG Pipeline: Retrieval vs. Generation (Interview Answer) · Bashiri Smith, Facebook · 2:23
- 5 Things AI Engineers Must Evaluate: RAG, Agents, Models, Data, Guardrails · Bashiri Smith, Facebook · 0:33
- Running LLM evals with promptfoo on local llama.cpp models · Melvin Vivas, X
- Cautionary Tale: Claude Code Wiped a Production Database via Terraform · Melvin Vivas, X
- Local PII Redaction on CPU with the AIBackends Python Package · Melvin Vivas, X · 0:14
Build
- Capstone v3, evals and guardrails: grow your 20 questions into a 40-case golden set that includes questions the documents can't answer and 10 prompt-injection cases (instructions hidden in a document, attempts to trigger the action tool). Score retrieval (hit@k) and answers (faithfulness and correctness with an LLM judge you spot-check against your own labels) in promptfoo, and add input/output screening plus PII redaction before anything is logged. (about 3.5 h) Idea from How to Evaluate a RAG Pipeline: Retrieval vs. Generation (Interview Answer).
Fill the gap (not from the creators)
- The lethal trifecta for AI agents (Simon Willison) · Article · simonwillison.net
A 10-minute mental model of why an agent that has private data, untrusted content and an outbound channel can be hijacked to leak data, and why guardrail filters alone don't fix it. (about 15 min) - Your AI Product Needs Evals (Hamel Husain) · Article · hamel.dev
The standard first read: look at your data, write assertion-style unit tests, add human and LLM-judge review, and let evals drive iteration, all shown on a real product. (about 30 min) - LLM Course: Evaluation (Maxime Labonne) · Course · github.com
Automated benchmarks, human evaluation and model-based (LLM-as-judge) evaluation, and why a metric that becomes the target stops being a good measure. (about 15 min) - LLM Course: Securing LLMs (Maxime Labonne) · Course · github.com
LLM-specific weaknesses (prompt injection, data and prompt leaking, jailbreaks, backdoors) and how to defend against them with red teaming and guardrails. (about 15 min)
Optional
- RAGAS · Tool · github.com
- Codex file deletions: why sandboxing matters for coding agents · Melvin Vivas, X
- AI Safety, Ethics, and Society Textbook (AI Safety Book) · Book · aisafetybook.com
Optional, to fill the gap (not from the creators)
- Stanford CME 295, Lecture 8: LLM Evaluation · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 8 of 9): LLM-as-a-judge: how to use it well, its benefits, and its biases and pitfalls. (about 2 h) - RAG Evaluation Using DeepEval & Confident AI (Yash Patil) · Video · youtube.com
Whole video: measuring a RAG pipeline with DeepEval metrics and tracking results in Confident AI. (about 29 min) - Defending LLM: Prompt Injection (LiveOverflow) · Video · youtube.com
Whole video: why LLMs are inherently open to prompt injection, and the design choices that limit the damage: treating input as untrusted, prompt design, user isolation, narrow tasks, few-shot prompts, fine-tuning, input limits, temperature 0 and redundant checks. (about 17 min)
Week 9: Ship, trace and control cost
Goal Deploy a containerized API with tracing, a fallback model, caching and a CI eval gate; report P50/P95 latency and cost per query. (about 9 h in total)
Core
- FastAPI · Tool · fastapi.tiangolo.com
Do: Deployment section only: 'Deployment concepts' and 'FastAPI in Containers - Docker' · about 45 min. - Langfuse · Tool · langfuse.com
Do: Get started, the Python/OpenTelemetry tracing quickstart, and the cost and latency views · about 1 h. - LiteLLM · Tool · github.com
Do: LiteLLM docs: the proxy Quickstart, Fallbacks (Reliability), Caching and Spend Tracking pages under docs.litellm.ai/docs/proxy/ · about 1 h.
From the creators' posts
- Taking a RAG App to Production: Evals, Guardrails, Cost and Tracing · Bashiri Smith, Facebook · 1:10
- Ollama vs vLLM: From Local AI Demo to Production Inference Serving · Bashiri Smith, Facebook · 0:45
- How to Reduce Latency in a Production AI Agent (Interview Answer) · Bashiri Smith, Facebook · 1:22
- RAG Pre-Deployment Checklist: Evals, Grounding, Latency, Cost & Monitoring · Bashiri Smith, Facebook · 0:54
Build
- Capstone v4, ship it: serve the agent with FastAPI in a Docker container on a free or cheap host, trace retrieval, model and tool calls in Langfuse, route model calls through LiteLLM with a fallback model and a cache, and make CI run your week-8 eval suite and block the deploy when scores drop. Record P50/P95 latency and cost per query, and use the production-evals chapter of the Evaluation Field Guide to decide what to monitor. (about 4.5 h) Idea from Taking a RAG App to Production: Evals, Guardrails, Cost and Tracing.
Fill the gap (not from the creators)
- Latency optimization guide (OpenAI API docs) · Docs · developers.openai.com
Seven provider-agnostic principles (process tokens faster, generate fewer tokens, use fewer input tokens, make fewer requests, parallelise, stream to cut perceived wait, don't default to an LLM) with a worked example. (about 30 min) - LLM Course: Deployment (Maxime Labonne) · Course · github.com
Ways to ship LLM apps, from local servers and demos to production servers and edge devices. (about 15 min) - LLM Course: Inference Optimization (Maxime Labonne) · Course · github.com
Flash Attention, the key-value cache and speculative decoding: the main ways to make generation faster and cheaper. (about 15 min) - Phase 17: Infrastructure and Production (AI Engineering from Scratch, Rohit Ghumare) · Course · github.com
Shipping LLM apps: managed platforms, model routing, gateways, monitoring and cost. Skim the lesson list and bookmark what your project needs. (about 15 min)
Optional
- How inference engines actually work (talk slides) · Article · x.com
- Large Language Model Operations (LLMOps) Explained (IBM Technology) · Video · youtube.com
- vLLM · Repo · github.com
Optional, to fill the gap (not from the creators)
- AI Gateways (AI Engineering from Scratch, Rohit Ghumare) · Tutorial · github.com
A lesson on gateways that sit between your app and model providers (LiteLLM, Portkey, Kong, Bifrost): routing, fallback, retries, rate limits and observability. (about 1 h) - Docker Tutorial for Beginners (TechWorld with Nana) · Course · youtube.com
Whole course: images, containers, Dockerfiles, Docker Compose, volumes and a full demo project. (about 3 h) - GitHub Actions Tutorial: CI/CD Pipeline with Docker (TechWorld with Nana) · Video · youtube.com
Whole video: workflows, events and jobs, and a CI pipeline that builds and pushes a Docker image. (about 32 min)
Week 10: Fine-tuning and system design
Goal Decide when to fine-tune, fine-tune a small router model and compare it with an LLM, and present your design trade-offs. (about 8.5 h in total)
Core
- Hugging Face LLM Course · Course · huggingface.co
Do: Chapter 11 only (chat templates, supervised fine-tuning with TRL, LoRA, evaluation) · about 2 h. - ModernBERT_train_classify.ipynb (donvito/notebooks) · Repo · github.com · the creator's own
Do: Run it end to end on a free Colab T4, then swap in your own labelled examples. For three classes, change num_labels and id2label/label2id in the model cell · about 1.5 h. - AI Agents vs AI Workflows: How to Choose for Production (AIBackends blog) · Article · aibackends.com · the creator's own
Do: Whole article (8 min): workflow vs. agent, the hybrid pattern and what production needs either way · about 15 min.
From the creators' posts
- Choosing a Knowledge Strategy: RAG vs Graph vs Fine-Tuning vs CAG vs Long Context · Bashiri Smith, Facebook · 0:23
- Model Distillation Explained: A Teacher Model Trains a Smaller Student · Melvin Vivas, X
- Train Your Own LLM Model Router by Fine-Tuning ModernBERT as a Classifier · Melvin Vivas, X · 1:00
- Production RAG Interview: Debugging Retrieval, Latency and Cost Like an Engineer · Bashiri Smith, Facebook · 1:47
Build
- Capstone v5, a fine-tuned router: label about 200 queries from your traces as 'small model is enough' or 'needs the strong model' (let your week-8 judge decide), fine-tune ModernBERT on them with the notebook, compare it with a zero-shot LLM classifier on accuracy, latency and cost, and let the winner pick the model for each /ask request. Then write a one-page design doc: architecture, failure modes, cost per query and why knowledge stays in RAG. (about 3.5 h) Idea from Train Your Own LLM Model Router by Fine-Tuning ModernBERT as a Classifier.
Fill the gap (not from the creators)
- Building A Generative AI Platform (Chip Huyen) · Article · huyenchip.com
A step-by-step reference architecture (context/RAG, guardrails, router and gateway, caching, write actions, observability, orchestration) that gives you the vocabulary design rounds expect. (about 1 h) - LLM Course: Supervised Fine-tuning (Maxime Labonne) · Course · github.com
Full fine-tuning compared with LoRA and QLoRA, the main training parameters, and the tools to run them; the next section covers preference alignment such as DPO. (about 15 min)
Optional
- Amazon Rufus (AI shopping assistant) · Article · amazon.science
- donvito/notebooks · Repo · github.com · the creator's own
- Zero-Shot Prompt Routing with Liquid AI's LFM 2.5-Encoder-350M · Melvin Vivas, X · 2:16
Optional, to fill the gap (not from the creators)
- Fine-Tuning with LoRA and QLoRA (AI Engineering from Scratch, Rohit Ghumare) · Tutorial · github.com
A hands-on lesson on why LoRA and QLoRA fit a 7B model in a few GB of memory, and how to set them up. (about 1.5 h) - Stanford CME 295, Lecture 4: LLM Training · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 4 of 9): pre-training, quantization, hardware optimization, supervised fine-tuning (SFT) and LoRA. (about 2 h) - Stanford CME 295, Lecture 5: LLM Tuning · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 5 of 9): preference tuning, RLHF, reward models, RL methods such as PPO, and DPO. (about 2 h) - System Design Concepts Course and Interview Prep (freeCodeCamp) · Video · youtube.com
Whole video: core system design concepts: scaling, load balancing, caching, databases and APIs. (about 53 min)
Week 11: Portfolio week
Goal Show a deployed, documented and evaluated project with a demo and a case study a hiring manager can scan in two minutes. (about 7.5 h in total)
Core
- How to Build a Coding Project That Will Get You Hired in 2026 [FULL BLUEPRINT] · Video · youtube.com · the creator's own
Do: Whole video (15 min) · about 15 min. - Bashiri Smith's AI Projects Guide · PDF · drive.google.com · the creator's own
Do: Open the project closest to yours; compare its architecture, phases and interview talking points with your capstone and pick one upgrade to add · about 30 min.
From the creators' posts
- Build Side Projects With Real Users to Stand Out to Employers · Bashiri Smith, Facebook · 0:59
- Stop Writing AI Keyword Soup: Build Projects That Solve Real Problems · Bashiri Smith, Facebook · 1:05
- Apply While You Build: Get Job-Search Feedback on Your AI Project Early · Bashiri Smith, Facebook · 1:12
Build
- Turn the capstone into a public case study: a README with the problem, users, architecture diagram, eval results (before and after hybrid search and reranking), latency and cost numbers, trade-offs and known limits; a 3-minute demo video; a live link; and a short write-up post. Ask 3-5 real users (teammates or friends) to try it and fix the top issue they hit. (about 6.5 h) Idea from Build Side Projects With Real Users to Stand Out to Employers.
Optional
- 27 AI Engineering Projects to Get Hired - The Ultimate Guide · Video · youtube.com · the creator's own
- AI Engineering Hub (patchy631/ai-engineering-hub) · Repo · github.com
Optional, to fill the gap (not from the creators)
- Project: Terminal Coding Agent (AI Engineering from Scratch, Rohit Ghumare) · Project · github.com
Capstone: build a terminal coding agent with a plan, a sandboxed tool surface and a loop that recovers from errors. A stretch project for after the plan. (about 35 h) - Project: Production RAG Chatbot (AI Engineering from Scratch, Rohit Ghumare) · Project · github.com
Capstone: a RAG chatbot for a regulated field, with hybrid search, reranking, citations and a golden-set, red-team and RAGAS evaluation. A stretch project for after the plan. (about 30 h) - Project: Multi-Agent Software Team (AI Engineering from Scratch, Rohit Ghumare) · Project · github.com
Capstone: an architect, parallel coders, a reviewer and a tester working as one team. A stretch project for after the plan. (about 40 h) - Project: GitHub Issue-to-PR Agent (AI Engineering from Scratch, Rohit Ghumare) · Project · github.com
Capstone: label an issue and get a tested pull request, built in a cloud sandbox with a minimal-permission GitHub App. A stretch project for after the plan. (about 30 h) - Project: MCP Server with Registry (AI Engineering from Scratch, Rohit Ghumare) · Project · github.com
Capstone: a stateless MCP server with tool discovery, a registry, authorization, policy and audit. A stretch project for after the plan. (about 25 h)
Week 12: Job search and trends
Goal Run a targeted job search, answer interview questions with numbers from your project, and explain where AI engineering roles are heading. (about 8 h in total)
Core
- The Interview Engine: The 7-Step System to Land AI Engineering Interviews · PDF · drive.google.com · the creator's own
Do: All 7 steps; set up the resume, the tracker and the outreach templates. The PDF has the prompts and email template; the tracker spreadsheet is only in the paid community, so make your own (Jobs, Warm paths, Contacts). The free resume template is linked from Bashiri's roadmap video · about 1.5 h. - 100 AI Engineering Interview Questions Guide · PDF · drive.google.com · the creator's own
Do: Skim all 100 questions (80 pages), mark the ones you can't answer yet, and practise 20 out loud · about 2 h. - The roles emerging in the gaps between roles (Andela Research) · Article · andela.com
Do: 'What the data shows' (three findings) and 'Reading skills, not titles' (10 min): AI engineering skills hide behind traditional job titles · about 15 min.
From the creators' posts
- Find Hidden AI Jobs: Search by Skills (RAG, LLMs), Not "AI Engineer" Titles · Bashiri Smith, Facebook · 1:03
- 90-Day AI Engineering Job Search: Targeted Companies, Outreach & Loom Pitches · Bashiri Smith, Facebook · 1:29
- AI Engineer Roadmap: 5 Skill Areas, a 24-Week Study Order & Interview Strategy · Bashiri Smith, Facebook · 2:29
- Coding agent harness knowledge as a new developer job skill · Melvin Vivas, X
Build
- Launch the job search: rewrite your resume as one page built around the capstone (problem, what you built, measured results), fill a tracker with 30 roles found by skill keywords (RAG, LLM, evals, vector database) rather than titles, send 10 tailored applications or referral requests with a short Loom walkthrough of your project, and do one mock interview using the 'Interview Preparation' section of the field guide from week 1. (about 4 h) Idea from 90-Day AI Engineering Job Search: Targeted Companies, Outreach & Loom Pitches.
Optional
- Why Your Job Applications Keep Getting Ignored · Video · youtube.com · the creator's own
- Latent Space · Newsletter · latent.space
- AI News Today (donvitocodes.com) · Website · donvitocodes.com · the creator's own
Optional, to fill the gap (not from the creators)
- Stanford CME 295, Lecture 9: Recap and Current Trends · Video · youtube.com
Free university lecture from Stanford Online (Autumn 2025, lecture 9 of 9): a recap of the whole course and the topics trending now. (about 2 h) - Forward Deployed Engineering 101 (Kevin Bai, Anthropic, at AI Engineer) · Video · youtube.com
Whole video: the forward-deployed model as Palantir ran it: selling an outcome rather than a product or a service, the business case for large contracts, building on a reusable platform instead of one-off custom work, two questions to ask before starting, and what has changed now that building is agentic. (about 18 min)