12-week study plan
The shortest sensible route through the whole guide, planned for about 8 to 12 hours a week. Each week has lessons and one practical task. Move on when the task is done, not when the calendar says so.
-
Week 1
The big picture and your setup
- Lessons
- Build
- Install Python and .NET, run the environment check, and write a one-page description of an AI idea at your workplace using the lifecycle steps.
-
Week 2
Data skills
- Lessons
- Build
- Clean and explore a messy synthetic customer table; create five features and explain each.
-
Week 3
How models learn
- Lessons
- Build
- Train a price predictor and a spam classifier; beat a baseline with both.
-
Week 4
Algorithms and evaluation
- Lessons
- Build
- Compare four algorithms with cross-validation; segment customers; choose a metric for a rare-event problem.
-
Week 5
Improve and apply ML
- Lessons
- Build
- Start capstone 1 (churn). Tune it, explain it and pick a threshold.
-
Week 6
Neural networks and vision
- Lessons
- Build
- Train a small network, watch for overfitting, and run a small image classifier.
-
Week 7
Transformers and other model families
- Lessons
- Build
- Run the attention demo and describe one business use for each model family.
-
Week 8
Inside a language model
- Lessons
- Build
- Train the tiny language model on your laptop and generate text at three temperatures.
-
Week 9
Scale, adapt and run LLMs
- Lessons
- Build
- Run a local model, measure its speed, and write a memory estimate for a model ten times larger.
-
Week 10
LLM apps and RAG
- Lessons
- Build
- Start capstone 4: a cited policy assistant with a 30-question evaluation set.
-
Week 11
Agents, quality and trust
- Lessons
- Build
- Add an approval-based tool to your assistant and run 15 adversarial tests.
-
Week 12
Production
- Lessons
- Build
- Serve capstone 1 as an API, add a promotion gate and a drift check, and draw its architecture.
After week 12. Finish at least two capstones (one ML, one AI application), then use the decision guide on the company cases to practise planning any new problem.