Data & AI
AI Application Development
This is product engineering where one component is probabilistic. The hard parts are not the model call: they are designing an interface that survives a wrong answer, and building automation that genuinely removes work rather than appearing to.
Start here
Pick a task you personally repeat every week. Automating something you actually resent doing is the difference between finishing this path and abandoning it at Stage 3.
Before you begin
- Can build a basic web application, or are learning to alongside this
Foundations
Enough programming and enough model understanding to build something real.
RequiredPython and working with APIs
Almost everything in this space runs through Python or TypeScript and a series of HTTP calls.
Python and working with APIs
Almost everything in this space runs through Python or TypeScript and a series of HTTP calls.
What to learn
- Python fundamentals
- Working with files and APIs
- JSON and data shapes
- Environment variables and secrets
- Virtual environments
Tools
- Python
- requests or httpx
Practice
Next: Understanding the model you are about to use.
RequiredWhat models can and cannot do
Building on a model without understanding its failure modes produces a product that breaks in ways you cannot explain to a user.
What models can and cannot do
Building on a model without understanding its failure modes produces a product that breaks in ways you cannot explain to a user.
What to learn
- Tokens and context limits
- Why models confabulate
- Cost per request
- Latency expectations
- Choosing a model for a task
Next: Automating something dull.
Automate something dull
Start with real work you personally repeat. Motivation matters more than novelty on this path.
RequiredWorkflow automation
Most valuable AI applications are ordinary automation with one intelligent step, not the other way round.
Workflow automation
Most valuable AI applications are ordinary automation with one intelligent step, not the other way round.
What to learn
- Scheduling
- Webhooks
- Third-party API integration
- Error handling and retries
- Notifications when something fails
Tools
- Python
- cron or a scheduler
- n8n or Zapier for the non-code path
Practice
Project
beginnerA personal automation
Automate a task you genuinely repeat: summarising a weekly report, sorting an inbox, extracting data from documents. It must run on a schedule without you.
- Python
- An LLM API
- A scheduler
Something you used to do by hand now runs on its own and you have stopped thinking about it.
Next: Making the intelligent step reliable.
Making the model step reliable
The difference between something that works when you demonstrate it and something that works.
RequiredPrompting and structured output
Free text cannot be consumed safely by the rest of your program.
Prompting and structured output
Free text cannot be consumed safely by the rest of your program.
What to learn
- Clear instructions and examples
- Structured output with a schema
- Validating every response
- Retrying on invalid output
- Versioning your prompts
Tools
- Pydantic or Zod
- Provider structured-output modes
Next: Grounding it in your own data.
RecommendedRetrieval over your own documents
The most common product request there is: 'can it answer questions about our documents?'
Retrieval over your own documents
The most common product request there is: 'can it answer questions about our documents?'
What to learn
- Embeddings
- Chunking
- Vector search
- Citing sources
- Saying 'I do not know' when the answer is absent
Tools
- An embedding model
- pgvector or a vector database
Next: The interface around it.
The product around the model
The part that decides whether people use it, and the part most engineers on this path underinvest in.
RequiredDesigning for a system that is sometimes wrong
The interface is where you decide what happens when the model is wrong, which it will be.
Designing for a system that is sometimes wrong
The interface is where you decide what happens when the model is wrong, which it will be.
What to learn
- Streaming to reduce perceived latency
- Showing sources and confidence
- Making output easy to correct rather than accept
- Never auto-applying a consequential action
- Setting expectations in the copy itself
Practice
Next: Building it as a real application.
RequiredBuilding and shipping the application
A script only you can run is not a product, and it teaches you nothing about how strangers behave.
Building and shipping the application
A script only you can run is not a product, and it teaches you nothing about how strangers behave.
What to learn
- A web interface
- Authentication
- Storing history and results
- Rate limiting and abuse
- Cost per user
Tools
- Next.js or FastAPI
- A database
- A hosting platform
Project
intermediateA tool other people use
A deployed application with accounts, a model-backed feature, saved history, cost limits per user, and a clear statement of what it cannot do.
- A web framework
- An LLM API
- A database
At least five people who are not you use it more than once.
Next: Judging whether it actually works.
Judging whether it works
The stage that separates a demo from something you would let an organisation depend on.
RequiredEvaluating outputs
'It seems good' is not a claim you can act on or defend. You need a number.
Evaluating outputs
'It seems good' is not a claim you can act on or defend. You need a number.
What to learn
- Building a test set from real usage
- Measuring accuracy on it
- Tracking quality after every prompt change
- Sampling outputs for human review
- Knowing your error rate
Practice
Next: Handling the errors you found.
RequiredHandling wrong answers and privacy
You are sending someone else's data to a third party and acting on probabilistic output. Both deserve care.
Handling wrong answers and privacy
You are sending someone else's data to a third party and acting on probabilistic output. Both deserve care.
What to learn
- Human review for consequential actions
- Graceful failure and escalation
- What data you send to a provider and why
- Retention and consent
- Prompt injection when input is untrusted
- Telling users what the system cannot do
Project
advancedA measured, honest product
Add an evaluation suite to your application that runs on every change, publish its accuracy in the interface, and add a human review path for anything consequential.
- An evaluation script
- Your application
You can tell a user how often the system is wrong and show them what happens when it is.
Where this leads
You do not have to pick one now. These are the directions this path opens up once you are working.
Continue your journey
The paths closest to this one. Skills overlap more than the job titles suggest.
AI Engineering
Build reliable systems on top of models you did not train.
ViewFull Stack Development
Own a feature from the database to the button someone clicks.
ViewProduct Management
Decide what gets built, why, and what gets cut.
ViewData Analysis
Turn messy data into a decision someone actually makes.
ViewYou do not have to do this alone
Our programs are free, taught live, and built around the same progression. Join one and work through it alongside other people.