Cracking ML Interviews

Open in the interactive study hub

2026 Interview Roadmap

A plan, not a reading list. This page tells you what is being tested in 2026, which topics matter for your specific role, and what to study in what order given how much time you have.

Not sure which role you're aiming for yet? Start with Choose Your Track, which picks a track from what you want to build and gives you a staged path through the repo. Come back here once you know the role, or when you're within a few weeks of interviewing.

Table of Contents

  1. What Changed for 2026
  2. Core Evaluation Areas
  3. Role-by-Role Focus Matrix
  4. The Typical Interview Loop
  5. Study Plans by Time Available
  6. Depth Checklist
  7. The Week Before
  8. Related Topics

What Changed for 2026

The classical ML fundamentals have not moved: bias/variance, metrics, regularization, and validation are still asked in nearly every loop. What changed is the layer on top.

Shift What it means for interviews
From "can you train a model" to "can you ship a system" Expect production, monitoring, and cost questions even in "modeling" rounds
Agents are the default architecture Tool use, planning loops, failure handling, and cost control are standard questions
RAG is assumed, not impressive You'll be asked why your retrieval failed, not what RAG is
Evaluation is the differentiator "How do you know it works?" is now the hardest question in most loops
Inference economics are interview material KV cache, batching, quantization, and cost-per-request appear in AI Engineer loops
Context engineering replaced prompt tricks Structuring what the model sees, not clever phrasing
Security is a real round Prompt injection, tool permissions, data exfiltration through agents

The single most common reason strong candidates fail an AI Engineer loop in 2026: they can build the happy path and cannot answer how they'd evaluate it, monitor it, or bound its cost.

Core Evaluation Areas

Every loop draws from these eight. The weight shifts by role, but none is safe to skip entirely.

# Area Sample of what's tested Start here
1 Classical ML fundamentals Bias/variance, regularization, cross-validation, imbalanced data, leakage Classical ML, Model Evaluation, Ensembles
2 Deep learning fundamentals Optimizers, normalization, attention, transfer learning, training failures Transformers, NN Training
3 LLM application engineering Prompting, structured outputs, tool use, RAG, context management LLM Fundamentals, RAG, Embeddings
4 Agents and orchestration Planning loops, tool schemas, multi-agent patterns, failure recovery, when not to use an agent Agentic AI, Multi-Agent Systems, MCP
5 Evaluation Offline vs online metrics, LLM-as-judge, regression suites, hallucination detection Model Evaluation, LLM Evaluation, Guardrails
6 MLOps and production Deployment, drift, experiment tracking, CI/CD, rollback, monitoring MLOps, Model Monitoring, Model Serving
7 System design Latency, throughput, cost, caching, batching, failure modes, capacity System Design, Design Patterns, Inference Optimization
8 Responsible AI and security Prompt injection, tool permissions, PII handling, bias, governance LLM Security, Explainability

Role-by-Role Focus Matrix

Priority: ●●● critical · ●● important · ● useful

Area ML Engineer AI / GenAI Engineer Data Scientist Data Engineer MLOps / Platform
Classical ML ●●● ● ●●● ● ●●
Statistics / experiment design ●● ● ●●● ● ●
Deep learning ●●● ●● ● - ●
LLM app engineering ●● ●●● ●● ● ●●
RAG and embeddings ●● ●●● ● ●● ●●
Agents and tool use ● ●●● ● - ●●
Inference optimization ●● ●●● - - ●●●
Evaluation ●●● ●●● ●●● ● ●●●
MLOps / deployment ●●● ●● ● ●● ●●●
Data pipelines / modeling ●● ●● ●● ●●● ●●
System design ●●● ●●● ● ●●● ●●●
SQL ●● ● ●●● ●●● ●●
Python / coding round ●●● ●●● ●● ●●● ●●
Docker / K8s / IaC ●● ●● - ●● ●●●
Security and governance ● ●●● ● ●● ●●

How to read this: your loop will test the ●●● rows properly and spot-check the ●● rows. Preparing the ● rows is what you do with leftover time, not first.

The Typical Interview Loop

Stage Format What determines the outcome
Recruiter screen 20-30 min Can you describe your work in plain language? Level calibration.
Technical screen 45-60 min Coding (Python/SQL) or an ML fundamentals rapid-fire round
Coding round 45-60 min Data manipulation, occasionally ML from scratch
ML/AI depth round 60 min Fundamentals with follow-ups going 3-4 layers deep
System design 60 min An open-ended "design X": the highest-variance round
Project deep-dive 45-60 min Your own past work, interrogated in detail
Behavioral / hiring manager 45 min Ownership, judgment, collaboration

Two rounds are under-prepared by almost everyone: the project deep-dive and system design. They are also the two with the widest score spread, which means they decide most loops. See Behavioral and Project Deep-Dive and the System Design framework.

Study Plans by Time Available

1 week: triage mode

Assume you cannot learn anything new; you can only surface what you already know.

Day Focus
1 Model Evaluation + your target role's ●●● rows, skimmed for gaps
2 Rehearse your project deep-dive out loud, twice. Look up your real numbers.
3 System design framework + one full case study end to end
4 Coding: Python or SQL, whichever your loop tests
5 Role-specific depth: RAG + Agents, or MLOps
6 2026 Additional Questions: answer out loud, don't read
7 Behavioral stories, rest, logistics

1 month: the realistic plan

Week Focus Deliverable
1 Fundamentals: classical ML, evaluation, ensembles, statistics Can explain bias/variance, leakage, and metric choice without notes
2 Deep learning + LLM: transformers, training, LLM fundamentals, embeddings, RAG A small working RAG app you can discuss
3 Production: MLOps, monitoring, serving, inference optimization, Docker Your project, containerized and monitored
4 Interview mechanics: system design cases, coding, behavioral, mock loops 3 rehearsed system designs, 8 STAR stories

3 months: build depth and a portfolio

Month Focus
1 Fundamentals with implementation: work through ML coding challenges, rebuild the classics from scratch, and read classical ML + deep learning properly
2 Build one real project end to end: ingestion, model or LLM pipeline, evals, API, container, monitoring. See Project Setup and Highlighted Projects. This project becomes your deep-dive story.
3 Interview preparation: system design breadth, 2026 questions, weekly mock interviews, behavioral story bank

The three-month plan works because month 2 gives you something real to talk about. A candidate with one deep project outperforms one with five tutorials, every time.

Depth Checklist

Use this to find gaps. If you cannot answer a line in two minutes without notes, that's your next study session.

Fundamentals

Deep learning

LLM and AI engineering

Production

Communication

The Week Before

View or edit this guide on GitHub