Common-Sense AI Engineering: Build LLM Apps & Agents from the Ground Up

Build core LLM app components yourself so you can see what frameworks usually hide.

Understand what LLMs are good at, where they fail, and why their behavior is fundamentally probabilistic.

Develop a mental model you can reuse across models, APIs, and frameworks.

Use prompts and context deliberately to shape behavior, reduce hallucinations, and improve consistency.

Learn when a bad result is really a context problem, not a model problem.

Reason about what the model sees, what it lacks, and what information it should get next.

Create semantic search with embeddings and vector storage before hiding retrieval behind a framework.

Connect LLMs to proprietary knowledge so answers are grounded in information the model did not train on.

Diagnose whether failures come from retrieval, context, or generation instead of blindly tweaking prompts.

Trace LLM behavior, inspect failures, and measure whether a change actually made your system better.

Turn recurring failure patterns into evals you can run again and again.

Improve quality systematically instead of relying on a handful of impressive examples.

Give an LLM tools that call real code, then build the control loop that lets it decide what to do next.

See exactly what makes an agent an agent before reaching for an agent framework.

Constrain agent behavior with explicit workflows when reliability matters more than autonomy.

Choose designs by balancing quality, latency, cost, reliability, and complexity, not by following hype.

Recognize when you need RAG, tools, workflows, a stronger model - or simply less machinery.

Leave with engineering judgment that transfers to the next generation of AI tools.