KolibríES

Technical Track

Courses for people already building with LLMs who need the system to survive production. Coding required.

Written without tying anything to a tool or model version: what is taught here still holds when the libraries change. Where something is not known with certainty, the course says so instead of filling the gap.

📐

Evaluating LLM Systems

Included in Kolibrí Pro

How to know whether your prompt change improved anything or broke something else: datasets, LLM-as-judge and its biases, and regression suites.

🎓 3 levels📚 12 lessons

1. Why You Need Evals

  • The problem: you do not know whether your change improved anything
  • Your evaluation set: where it comes from and how to build it
  • What can actually be measured, and what cannot
  • The small-numbers mistake

2. LLM as a Judge

  • LLM as a judge: how it works and why it is so widely used
  • The judge's biases, and what to do about each
  • Calibrating the judge against humans
  • Rubrics that produce agreement

3. Evals in Your Workflow

  • The regression suite: making sure a fix does not break what worked
  • Offline and online: two different questions
  • Traces: without observability there is no diagnosis
  • Your minimum eval system, running this week
🔍

RAG That Works: Retrieval, Not Magic

Included in Kolibrí Pro

Why your document chatbot answers badly: chunking, the real limits of embeddings, hybrid search, reranking, and how to measure each step.

🎓 3 levels📚 12 lessons

1. Why Retrieval Fails

  • RAG is search before it is generation
  • Chunking: the decision with the most impact and the least thought
  • Embeddings: what they capture and what they do not
  • Measuring retrieval on its own

2. Retrieving Better

  • Hybrid search: why BM25 is still alive
  • Reranking: the step almost nobody adds
  • Metadata and filtering: the cheapest improvement
  • The user's question is not the search query

3. RAG in Production

  • Lost in the middle: context order matters
  • Citing, and knowing how to say you do not know
  • The index changes: maintenance and data that expires
  • When RAG is not the answer
🛡️

Securing LLM Applications

Included in Kolibrí Pro

Direct and indirect prompt injection, over-privileged tools and data exfiltration: why there is no fix, and how to shrink the blast radius.

🎓 3 levels📚 12 lessons

1. The Underlying Problem

  • There is no separation between instructions and data
  • Direct and indirect injection: the second is the dangerous one
  • Blast radius: the only variable you control
  • OWASP for LLM as a map, not a certificate

2. Attack Surfaces

  • Tools: where injection becomes an incident
  • Poisoned RAG: when the attacker writes your context
  • Data and system-prompt leakage
  • Cost and availability as an attack vector

3. Defence in Depth

  • Least privilege: the defence that actually works
  • Model output is untrusted input
  • Human in the loop, where it actually matters
  • Testing your own system before someone else does