Become an
ML engineer
in 2026

• Pre-training architecture research (transformers, diffusion, world models).
• Distributed training on thousands of TPUs and GPUs.
• RLHF pipeline engineering.
• Scaling law experiments and efficiency optimisation.
• Building evaluation frameworks.

• Extremely competitive. PhD often expected.
• Small teams. Roughly 100 to 200 core researchers per lab.
• Training runs cost $10M to $100M+.
• Long cycles per run.

• Strong ML fundamentals plus published research.
• Exceptional distributed systems engineering.
• Domain expertise, e.g. CV to multimodal.

• Fine-tuning foundation models for specific domains.
• Building RAG pipelines.
• Deploying models with monitoring.
• Building multi-agent systems and agentic workflows.
• Optimising inference cost and latency.
• A/B testing model performance in production.

• Accessible entry. Bootcamps plus portfolio work is the path.
• Rapid skill obsolescence. Tools shift fast.
• Need to ship fast. MVP in a stage, not a semester.
• Thousands of open roles globally.

• Strong SWE plus a portfolio of deployed AI projects.
• Deep knowledge of one cloud platform.
• Two to three end-to-end projects, GitHub plus live demos.
• Open source contributions.

• Building feature stores and model registries.
• Creating self-service ML platforms.
• Optimising model serving for latency, throughput, cost.
• Managing GPU and TPU clusters and orchestration.
• CI/CD pipelines for ML workflows.
• Model monitoring and observability.

• Usually needs a couple of cycles of SWE experience first.
• Bridges ML plus DevOps. Rare, so high value.
• Less about models, more about reliable systems.
• Critical for scaling ML in organisations.

• SWE, then learn ML fundamentals, then specialise in MLOps.
• SRE background plus ML interest.
• Strong DevOps or platform engineering plus ML projects.

Prompt engineering

System prompts, chain-of-thought, structured outputs, function calling.

RAG systems

Chunking strategies, vector DBs, hybrid search, retrieval evaluation.

Fine-tuning

LoRA / QLoRA, dataset prep, when to fine-tune vs RAG vs prompting.

Multi-agent systems

ReAct, Plan-and-Execute, LangGraph, tool use orchestration.

The differentiator. Separates hobbyists from professionals.

Deployment

Docker, FastAPI, serverless vs container, model serving (vLLM).

Monitoring

Accuracy drift, data drift detection, A/B testing frameworks.

Cost optimisation

$/token analysis, caching, quantisation (8-bit, 4-bit), batching.

Cloud platforms

Pick one deeply: AWS SageMaker, GCP Vertex AI, or Azure ML.

• Endless theory before building anything.
• Following a linear roadmap: math, then basics, then DL, then prod.
• Collecting certificates instead of building projects.
• Learning every framework shallowly.
• Waiting to "feel ready" before deploying.

• You learn by building, breaking, and fixing.
• Skills are learned in parallel, not sequence.
• Your GitHub is more valuable than your resume.
• Deployed demos beat polished code nobody can access.
• You are never "ready". You ship, and then you learn.

• Drop your current tier in chat. Foundation, Applied, or Platform?
• One line. Not the "why". Just the tier.
• Three or four get read aloud.
• Then straight into the portfolio-first path.

Most attendees cluster on Tier 2

That is the largest market, and the path we cover in most depth from here. Tier 1 and Tier 3 attendees get pointed at the specific extras they need in the resources block near the close.

The next block is the phased path: what to build, in what order, with a deployed portfolio piece at every stage.

• Fast.ai · Practical Deep Learning for Coders (free).
• DeepLearning.AI · Short courses on LangChain, agents.
• HuggingFace · Transformers Course.
• Full Stack Deep Learning · Production focus.
• Andrej Karpathy · Neural Nets: Zero to Hero.

• Papers: Arxiv Sanity, Papers with Code.
• Newsletters: The Batch, Import AI.
• Podcasts: Latent Space, TWIML.
• Communities: r/ML, HuggingFace Discord.