Choose Your Track

This repository has over 100 guides. You do not need all of them: you need one track, in order.

Pick a track below, follow its stages top to bottom, and skip everything else until you finish. Coming back for a second track later is much easier than trying to learn two at once.

Table of Contents

1. The 30-Second Chooser
2. Pick by Where You're Coming From
3. The Common Core
4. Track 1: Machine Learning Engineer
5. Track 2: AI / GenAI Engineer
6. Track 3: Data Scientist
7. Track 4: Data Engineer
8. Track 5: MLOps / Platform Engineer
9. Track 6: Deep Learning Engineer
10. Track Comparison
11. How to Actually Study This
12. Frequently Asked
13. Related Topics

The 30-Second Chooser

Find the row that sounds most like you.

Still unsure? Two honest tie-breakers:

• The most in-demand entry point right now is the AI / GenAI track. It has the shortest path from zero to something working, and it needs the least mathematics. If you want to be employable fastest, start there.
• The most durable foundation is the ML Engineer track. It is slower, but everything else becomes easier afterwards, and it does not go stale when the tooling changes.

You can also just start with the Common Core below: the first two weeks are identical for every track, so you can defer the decision.

Pick by Where You're Coming From

The Common Core

Every track needs these. If you are early, do this first and decide afterwards: roughly 1-2 weeks.

1. Python for AI Engineering: async, typing, Pydantic, APIs, logging
2. Statistics & Probability: distributions, hypothesis testing, the CLT
3. Model Evaluation and Metrics: the single most reusable guide in this repo
4. Python Coding Challenges and SQL Coding Challenges
5. GitHub Project Setup, so your work is shareable from day one

Everything after this is track-specific.

Track 1: Machine Learning Engineer

Job titles: Machine Learning Engineer, Applied Scientist, ML Software Engineer Prerequisites: comfortable Python; basic linear algebra and probability helps but can be learned alongside Realistic time to interview-ready: 3-4 months from the Common Core, part-time

Stage 1: Foundations (3-4 weeks)

• Classical ML Overview: algorithms, when each applies
• Model Evaluation and Metrics
• Feature Engineering & Selection
• Ensemble Methods and Gradient Boosting: what actually wins on tabular data
• Clustering and Dimensionality Reduction

Stage 2: Deep learning (3-4 weeks)

• Deep Learning Overview
• Neural Network Training: optimizers, normalization, debugging
• Sequence Models → Transformers
• PyTorch

Stage 3: Production (3-4 weeks)

• MLOps Overview
• Model Serving and Model Monitoring
• CI/CD for Machine Learning
• MLflow · Docker

Stage 4: Interview prep (2-3 weeks)

• ML System Design Framework + both case studies
• ML Coding Challenges
• 2026 Interview Roadmap · Behavioral Guide · Take-Home Projects

Build this: a tabular prediction model served behind a FastAPI endpoint, containerized, with tracked experiments and a monitoring dashboard. End to end matters more than accuracy.

You're ready when you can explain bias/variance without notes, choose a metric and defend it against alternatives, spot data leakage in someone else's setup, and describe how you would monitor and roll back a deployed model.

Track 2: AI / GenAI Engineer

Job titles: AI Engineer, GenAI Engineer, LLM Engineer, Applied AI Engineer Prerequisites: solid Python and API work. You do not need to know how to train a model. Realistic time to interview-ready: 2-3 months, part-time: the shortest path here

Stage 1: LLM foundations (2-3 weeks)

• LLM & Generative AI Fundamentals
• Prompt Engineering
• Context Engineering: what actually goes in the window, and what it costs
• Structured Outputs & Function Calling

Stage 2: Retrieval (2-3 weeks)

• Intro to RAG → RAG Engineering
• Embeddings: chunking and retrieval quality live here
• Vector Databases → Advanced

Stage 3: Agents (2-3 weeks)

• Agentic AI → Agent Systems & Tool Use
• Multi-Agent Systems
• MCP Protocol · LangGraph

Stage 4: Production (2-3 weeks)

• LLM Evaluation and Evaluation & Guardrails: the most-tested area in these interviews
• LLM Inference Optimization: cost and latency
• LLM Security: prompt injection is a real round
• LLMOps · FastAPI · Docker

Stage 5: Interview prep (2 weeks)

• Backend AI System Design and ML System Design Patterns
• 2026 Interview Questions · Behavioral Guide

Build this: a RAG assistant over a real document set with citations, hybrid search, a reranker, an evaluation harness, and a cost-per-request number you can quote. The eval harness is what separates candidates.

You're ready when you can diagnose whether a bad RAG answer came from retrieval or generation, design an eval suite for a non-deterministic feature, explain prompt injection and how tool permissions bound it, and state your system's cost per request and the three biggest levers on it.

The deeper curriculum for this track is The Agentic AI Engineer Roadmap: 34 topics from foundation to frontier. Use it after Stage 3.

Track 3: Data Scientist

Job titles: Data Scientist, Product Data Scientist, Quantitative Analyst Prerequisites: SQL, and genuine comfort with statistics Realistic time to interview-ready: 2-3 months, part-time

Stage 1: Statistical foundation (3-4 weeks)

• Statistics & Probability: go deep; this is the whole round
• Model Evaluation and Metrics
• SQL Coding Challenges: window functions until they're automatic

Stage 2: Modeling (3 weeks)

• Classical ML Overview
• Feature Engineering · Ensemble Methods
• Time Series · Clustering

Stage 3: Experimentation and communication (2-3 weeks)

• A/B Testing: the defining skill of this role
• Model Explainability: SHAP for stakeholder conversations
• Data Quality

Stage 4: Interview prep (2 weeks)

• Take-Home Projects: data science loops lean on these heavily
• Behavioral Guide · Python Coding Challenges

Build this: an end-to-end analysis answering a real business question: framing, EDA, a model, an experiment design, and a written recommendation. The write-up is the deliverable, not the notebook.

You're ready when you can design an A/B test including sample size and guardrails, explain a model's output to a non-technical stakeholder in their vocabulary, write window-function SQL fluently, and recognize Simpson's paradox and the peeking problem in a described scenario.

Track 4: Data Engineer

Job titles: Data Engineer, Analytics Engineer, Data Platform Engineer Prerequisites: strong SQL, comfortable Python Realistic time to interview-ready: 2-3 months, part-time

Stage 1: Modeling and architecture (2-3 weeks)

• Data Modeling: star schemas, slowly changing dimensions
• Data Architecture: warehouse, lake, lakehouse
• SQL Coding Challenges

Stage 2: Pipelines (3 weeks)

• Data Processing Pipelines: batch vs streaming, and when each is right
• Apache Spark · Apache Airflow
• dbt → dbt Interview Q&A

Stage 3: Modern stack (2-3 weeks)

• Apache Kafka: streaming
• Delta Lake · Apache Iceberg · DuckDB
• Data Quality

Stage 4: Serving ML (1-2 weeks)

• Data Engineering for AI: point-in-time correctness, training/serving skew
• Feature Stores

Stage 5: Interview prep (2 weeks)

• Backend System Design Interview Guide
• Docker · Behavioral Guide

Build this: an ingestion → transform → warehouse pipeline orchestrated in Airflow, modeled in dbt, with data quality tests that block on failure and a documented lineage graph.

You're ready when you can design a star schema from requirements, explain idempotency and why partial writes must never be visible, argue batch versus streaming from the latency the business actually needs, and describe how late-arriving data is handled.

Track 5: MLOps / Platform Engineer

Job titles: MLOps Engineer, ML Platform Engineer, ML Infrastructure Engineer Prerequisites: real software engineering experience; Linux and containers Realistic time to interview-ready: 2-3 months if you already do DevOps

Stage 1: ML literacy (2 weeks)

You need enough to reason about models, not to build them.

• Classical ML Overview · Model Evaluation

Stage 2: The MLOps core (3-4 weeks)

• MLOps Overview
• CI/CD for Machine Learning: registry, gates, canary, rollback
• Model Serving · Model Monitoring
• MLflow · Feature Stores · Data Quality

Stage 3: Infrastructure (2-3 weeks)

• Docker → Kubernetes → Helm
• Terraform · GitHub Actions
• Cloud ML Platforms, then your employer's cloud

Stage 4: LLM operations (2 weeks)

• LLMOps · LLM Inference Optimization
• LLM Evaluation · Testing AI Systems

Stage 5: Interview prep (2 weeks)

• ML System Design · Backend AI Systems
• A/B Testing · Behavioral Guide

Build this: a deployment pipeline that trains, gates on evaluation, registers, deploys to a canary, and rolls back automatically when metrics degrade. The rollback is the interesting part.

You're ready when you can explain training/serving skew and how a feature store prevents it, describe a complete rollback (model and feature pipeline), design promotion gates that catch a slice regression, and estimate GPU cost per thousand predictions.

Track 6: Deep Learning Engineer

Job titles: Deep Learning Engineer, Research Engineer, Applied Scientist Prerequisites: linear algebra, calculus, probability. This track is the most mathematically demanding. Realistic time to interview-ready: 4-5 months, part-time

Stage 1: Fundamentals (4 weeks)

• Deep Learning Overview
• Neural Network Training: read this twice
• PyTorch
• ML Coding Challenges: implement backprop and attention from scratch

Stage 2: Architectures (4 weeks)

• Sequence Models: why gating works, and what attention replaced
• Transformers: the central topic
• Computer Vision
• Applied Deep Learning

Stage 3: Modern practice (3 weeks)

• Fine-Tuning: LoRA, QLoRA, PEFT, DPO
• HuggingFace · Unsloth
• LLM Inference Optimization · vLLM

Stage 4: Interview prep (2 weeks)

• ML System Design · 2026 Interview Questions
• Behavioral Guide

Build this: a transformer implemented from scratch, then a fine-tune of an open model on a domain dataset with a proper evaluation and a documented ablation.

You're ready when you can derive attention including why scores are scaled by √d_k, diagnose a NaN loss in order of likelihood, explain what to do when a model won't fit in GPU memory in escalating order, and describe LoRA's mechanism and memory savings.

Track Comparison

How to Actually Study This

Follow the stages in order. They are ordered by dependency, not by interest. Skipping ahead to agents before understanding evaluation is the most common way people stall.

Build after every stage. Reading a guide creates recognition, not recall. A small project after each stage converts one into the other, and it gives you the material for the project deep-dive round, which decides more interviews than the technical rounds do.

Answer out loud. Every guide ends with an Interview Q&A section. Read the question, answer it aloud before reading on, and notice the gap. Silent reading consistently overestimates how well you know something.

One track at a time. Two half-finished tracks interview worse than one finished one.

Depth beats breadth in interviews. One project you can discuss five layers deep beats five tutorials. Interviewers probe until they find your limit; the goal is for that limit to be deep, not wide.

When you're within a few weeks of interviewing, switch to the 2026 Interview Roadmap: it has the role-by-role focus matrix and study plans for 1 week, 1 month, and 3 months.

Frequently Asked

Do I need a PhD? No, for every track here except some research positions. PhDs matter for research scientist roles at frontier labs. ML Engineer, AI Engineer, Data Engineer, and MLOps roles are engineering roles that hire on demonstrated ability.

How much maths do I actually need? For AI/GenAI, Data Engineering, and MLOps: very little beyond comfort with probability and reading a formula. For ML Engineer: linear algebra and probability at an undergraduate level. For Data Science: real statistical depth, since it is the interview. For Deep Learning: solid linear algebra, calculus, and probability.

Can I switch tracks later? Yes, and it gets easier each time: the Common Core is shared, and adjacent tracks overlap heavily. ML Engineer ↔ Deep Learning and AI Engineer ↔ MLOps are the easiest moves.

Do I have to read every guide in my track? No. The stages are the path; the rest of the directory is reference. Use the "Guides in This Track" index in each directory's README when you need something specific.

How long does this really take? The estimates assume 10-15 hours a week and existing programming ability. Full-time roughly halves them. Starting from no programming background, add 2-3 months for Python first.

Is the AI/GenAI track a shortcut? It is faster to a first job, and that is real. It is also a shallower foundation: if the tooling shifts, ML Engineer fundamentals transfer and prompt-level knowledge does not. Many people start with AI/GenAI for the job and backfill fundamentals afterwards, which is a perfectly reasonable order.

I already work in ML. Where do I start? Skip to the stage covering what you don't do daily. Most working MLEs have gaps in evaluation, monitoring, and system design rather than in modeling.

Related Topics

• 2026 Interview Roadmap: role focus matrix and time-boxed study plans
• Study Pattern: topics by difficulty level
• Behavioral and Project Deep-Dive Guide
• ML Take-Home Projects
• 2026 Additional Questions · 2026 Interview Questions
• Resources and References
• Repository Home