AI Engineer
Build products on top of foundation models.
You want to build with large language models, not train them.
- 01Build retrieval pipelines that answer from your own data, with citations
- 02Ship agents that use tools and recover when a tool fails
- 03Prove a change helped, using evals, tracing and error analysis
Your AI Engineer career begins
AI engineering is not machine learning with a new name. You are not collecting datasets or training networks — you are taking models that already exist and turning them into features that behave predictably enough to put in front of users.
That turns out to be its own discipline. Retrieval decides your quality ceiling long before prompting does. Agents fail in ways ordinary code does not. And the skill that separates people who ship from people who demo is evaluation: knowing whether a change made the system better, rather than believing it did.
This track goes in that order — foundations, retrieval, agents, evaluation, then the production concerns that keep an AI feature alive: tracing, cost, caching, guardrails, and defending against inputs designed to break it.
Not model training or research — that is the ML Engineer track.
The AI Engineer job
- Turn foundation models into product features that behave predictably
- Design and tune retrieval pipelines over company data
- Build agents that use tools and recover from failure
- Own evaluation, tracing, cost, and the safety of what ships
- Product companies adding AI to something people already use
- AI-native startups where the model is the product
- Enterprise teams building internal assistants over their own data
- Consultancies delivering AI systems for clients
- 1Junior AI Engineer — Builds and tunes features against existing pipelines.
- 2AI / GenAI Engineer — Owns a retrieval or agent system end to end.
- 3Senior AI Engineer — Sets evaluation standards and architecture.
- 4Staff / Lead AI Engineer — Owns the AI platform other teams build on.
Every skill in the track, in the order you learn it.
6 phases, 116 named skills. Not chapter headings — the individual things you will be able to do by the end.
- PHASE 1 OF 620 SKILLS
Python and model foundations
The language, and what is actually happening inside an API call.
- Python typing and dataclasses
- Comprehensions
- Virtual environments
- Async Python
- Concurrent API calls
- HTTP and REST clients
- Tokens and tokenisation
- Context windows
- Transformer intuition
- Attention at working depth
- Temperature and top-p
- Sampling behaviour
- System and user roles
- Few-shot prompting
- Chain-of-thought
- Reasoning models
- Structured outputs
- JSON schema enforcement
- Model trade-offs
- Rate limits and latency
- PHASE 2 OF 624 SKILLS
Retrieval and RAG
Where most AI products are actually won or lost.
- Embeddings
- Cosine similarity
- Distance metrics
- Fixed-size chunking
- Recursive chunking
- Semantic chunking
- Hierarchical chunking
- Overlap tuning
- PDF and HTML parsing
- Table and code extraction
- Metadata design
- Filtered retrieval
- pgvector
- Pinecone, Weaviate and Qdrant
- BM25 and TF-IDF
- Hybrid search
- Score fusion
- Cross-encoder reranking
- Query rewriting
- Query decomposition
- Multi-hop retrieval
- Grounding and citations
- recall@k, MRR and nDCG
- Knowing when RAG is wrong
- PHASE 3 OF 619 SKILLS
Agents and tool use
Giving a model the ability to act, and containing what happens next.
- Function calling
- Tool schema design
- The reason-act loop
- Planning vs execution
- Model Context Protocol
- Sandboxed code execution
- Web search tools
- Retries and timeouts
- Failure recovery
- Short-term memory
- Long-term memory
- Context engineering
- Context compaction
- Planner-executor patterns
- Reviewer agents
- Agent handoffs
- Shared state
- Graph orchestration
- Human-in-the-loop
- PHASE 4 OF 617 SKILLS
Evaluation and error analysis
The difference between shipping and guessing.
- Golden datasets
- Task-specific metrics
- Schema and regex checks
- Assertion suites
- LLM-as-judge
- Rubric design
- Judge calibration
- Pairwise comparison
- Preference data
- Retrieval vs generation evals
- Faithfulness metrics
- Error analysis loops
- CI regression suites
- Offline vs online eval
- Prompt A/B testing
- Annotation workflows
- Inter-rater agreement
- PHASE 5 OF 620 SKILLS
Production and reliability
Everything that stands between a working demo and a live feature.
- OpenTelemetry tracing
- Span-level debugging
- Prompt and completion logging
- Latency budgets
- Response streaming
- Prompt caching
- Semantic caching
- Request batching
- Backoff and queueing
- Fallback models
- Degraded modes
- Cost attribution
- Prompt versioning
- Safe rollout
- Input guardrails
- Output guardrails
- PII detection and redaction
- Prompt injection
- Jailbreak defence
- Data exfiltration risks
- PHASE 6 OF 616 SKILLS
Beyond the API
When calling a hosted model stops being enough.
- Prompting vs RAG vs fine-tuning
- Supervised fine-tuning
- Dataset curation
- LoRA
- QLoRA
- Model distillation
- Quantisation trade-offs
- Open-weight models
- Self-hosting with vLLM
- Running models with Ollama
- GPU memory constraints
- Vision inputs
- Audio and speech agents
- Voice pipelines
- Structured extraction at scale
- Synthetic data generation
6 projects. Every one of them a running system.
Not exercises with a known answer. Each project is something you can deploy, send someone a link to, and defend in an interview.
A prompt workbench with structured outputs
A small tool for running the same prompt across models and settings, enforcing a JSON schema on the result, and tracking what each run cost.
Semantic search over a document collection
Parse, chunk, and embed a real corpus, then search it by meaning rather than keywords — and measure how often the right passage comes back.
A production RAG assistant with citations
Hybrid retrieval, reranking, query rewriting, and answers grounded in sources the user can click — plus the evaluation set that proves each change helped.
A tool-calling agent that completes real tasks
An agent with a real toolset, a reason-act loop you wrote yourself, memory across turns, and sane behaviour when a tool fails or returns nonsense.
An evaluation harness running in CI
A golden dataset, programmatic checks, a calibrated LLM judge, and a regression suite that fails the build when a prompt change makes things worse.
A multi-agent system with tracing and guardrails
Several agents coordinating on a real workflow, fully traced span by span, with input and output guardrails, cost caps, and defences against injected instructions.
AI Engineer · GenAI Engineer · LLM Application Engineer
From first lesson to signed offer.
Learn
Work through the curriculum in order, from the fundamentals to the applied work the role does every day. Written lessons, not video you cannot search.
Build
Ship the projects. Each one is a real system with real constraints, and each one ends up somewhere you can show it.
Practice
Solve problems in the DSA arena, take AI mock interviews, and get feedback on the answers you actually gave.
Get hired
Build an ATS-ready resume, see how well you match each opening, close the gaps it finds, and apply without leaving the platform.
Questions people ask before starting.
Do I need a machine learning background?
No. You need to be comfortable programming and to understand models well enough to use them — tokens, context, embeddings, sampling. The maths-heavy path of training your own models is the ML Engineer track.
Is this just prompt engineering?
No. Prompting is one phase of six. The work that makes the role hard is retrieval quality, agent reliability, and evaluation — knowing whether a change actually improved the system rather than assuming it did.
Which frameworks does the track use?
You build the core loops by hand first, so you understand what a framework is doing, then work with the common orchestration and evaluation tooling. The point is the concepts, which outlive whichever library is current.
Will this still be relevant when the models change?
The models change constantly; the engineering around them moves much more slowly. Retrieval design, evaluation discipline, tracing, cost control, and injection defence apply no matter which model you call.
Do I need a GPU?
Not for most of the track — you work against hosted APIs. The final phase covers self-hosting and fine-tuning, and explains what hardware each option needs.
The other three tracks.
Full Stack AI Engineer
Most popularShip the entire product — interface, API, and the AI layer.
You want to build and launch complete AI products on your own.
The superset track: everything an AI Engineer does, plus the product around it.
You will build- A responsive marketing site with a real content source
- A REST API with authentication and PostgreSQL
- A real-time collaborative board
Roles: Full Stack Engineer · Product Engineer (AI) · Founding Engineer
Explore this track→ML Engineer
Train, serve, and monitor models in production.
You want to own models end to end, from raw dataset to live endpoint.
Not calling somebody else’s API — here you build and own the model.
You will build- An end-to-end tabular model with an honest evaluation
- An image classifier fine-tuned from a pretrained backbone
- A reproducible training pipeline with tracked experiments
Roles: Machine Learning Engineer · MLOps Engineer · Applied ML Engineer
Explore this track→Data Engineer
Move and model data at scale.
You want to build the pipelines every other team depends on.
Not dashboards and analysis — you build the systems those run on.
You will build- An analytics layer over a raw dataset
- An incremental ingestion job with tests
- A batch ELT pipeline: orchestration, transformation, warehouse
Roles: Data Engineer · Analytics Engineer · Data Platform Engineer
Explore this track→Everything you need to grow as an engineer — software engineering, AI, real-world projects, technical challenges, hands-on practice, AI-powered guidance, interview preparation, and career development.
Don’t just learn engineering. Experience it.
Learn. Build. Solve. Practice. Prepare. Get hired.