Generative AI & LLM Engineering Foundations
Learn to reason about model internals, adaptation budgets, evaluation, tool approval, and release operations through an offline support-category workbench.
What you'll be able to do
- Explain tokenization, attention, and supervised loss with small calculations
- Design leakage-resistant datasets and measured adaptation experiments
- Separate model suggestions from authorized actions
- Release a versioned model system with privacy-aware telemetry and rollback evidence
Before you start
- Comfortable with Python, basic algebra, probability, and test design
- Understand training versus inference and can read a confusion matrix
- Python 3.11 or newer; optional model training requires a separately reviewed hardware and dependency plan
The curriculum
- From Tokens to Context — Free preview
Use a tiny attention calculation to distinguish model representations from understanding, memory, and truth.
- Datasets That Can Disagree With You — Free preview
Split related examples together, define loss targets deliberately, and prevent evaluation leakage.
- Adapt With a Measured Budget — Sign-in access
Understand low-rank updates and quantization costs before choosing optional hardware-heavy experiments.
- Make Baselines Hard to Beat — Sign-in access
Compare candidates using category-level errors, abstentions, and a frozen evaluation protocol.
- Proposals Before Actions — Free preview
Keep model-generated suggestions separate from authorized tool execution and durable workflow state.
- Latency Has Components — Sign-in access
Measure request stages and design caches that respect versions, identity, and changing access rules.
- Observe Without Exposing — Sign-in access
Use telemetry and rollout gates to detect failures while limiting unnecessary collection of user content.
- Release Signal Bench — Sign-in access
Integrate the offline classifier, review workflow, measurements, and compatible rollback into a demonstrable capstone.
Signal Bench
Build a local release workbench that classifies fictional support messages into categories, compares a baseline with a candidate, routes ambiguous cases to review, and records reproducible release decisions.
Deliverables
- Synthetic grouped dataset and frozen evaluation set
- Offline baseline classifier and candidate comparison
- Reviewed action proposal state machine
- Latency measurements, release manifest, and rollback demonstration
Review your work
- Related examples do not cross training and evaluation groups
- Category-level errors and abstentions are visible rather than hidden by aggregate accuracy
- Model output cannot directly trigger an external action
- Release gates and rollback are demonstrated using synthetic fixtures without paid calls