Interactive labs · 3D · free
Learn AI engineering by fixing broken systems
Each lab drops you into a production incident: a RAG pipeline citing the wrong text, an agent looping forever, a guardrail that fails open. You see the system in 3D, find what's wrong and fix it with the same controls engineers use. About five minutes each, in your browser, no setup.
0 of 12 labs fixed in this browser
01
Read the brief
What broke in production, and what fixed looks like.
02
Watch it work
Step through the system in 3D, one part at a time.
03
Spot the fault
Find where it breaks before you touch anything.
04
Fix it
Tune real controls until every target turns green.
All labs
In curriculum order. Start anywhere: each lab stands on its own.
Fix a search that can't find the right passage
Support search returns the wrong policy. Queries and documents were embedded differently, so the query lands far from the passages that answer it.
Make vector search fast without losing recall
The vector store scans every chunk for every query, so search is far too slow, and the permission filter runs after the search and throws away results.
Stop a RAG pipeline citing the wrong text
Answers cite text that doesn't contain the answer. Chunking cuts facts in half, and vector-only search misses exact policy IDs.
Pack a context window instead of stuffing it
A long-running assistant sends every turn and every retrieved chunk to the model. The window overflows, the key fact sits in the middle where attention is weakest, and the answer gets cut off.
Get reliable JSON out of a model
An extraction endpoint uses plain JSON mode at a high temperature. Replies parse as JSON but break the schema, and refusals crash the parser.
Follow one tool call and make it safe
A weather question gets routed to a tool that deletes records. Tool descriptions are vague, arguments aren't validated and a slow API can hang the whole request.
Stop an agent looping forever
An agent keeps calling the same tool with reworded arguments. Nothing stops it, so it never answers and the bill keeps growing.
Lock down an MCP setup
An agent host is plugged into three MCP servers. The database server can write, a community server floats on its latest version, and an injected tool result reaches the database.
Catch a regression before users see it
Releases that make answers less faithful keep reaching users. The judge is the same model as the system it grades, and nothing blocks a bad release in CI.
Line up guardrails so attacks can't get through
Only an input filter protects the assistant, and when it times out it lets everything through. Injected instructions in a retrieved document and leaked secrets pass straight to users.
Find the slow span in a trace
Users wait seconds for a first word. Steps that could run together run one after another, nothing streams, and the trace store throws away the one error trace you need.
Fine-tune on one GPU without breaking the model
A team is fully fine-tuning an 8B model on noisy data to teach it facts. It doesn't fit in GPU memory and it forgets what it could already do.
Prefer writing code? The hands-on coding labs → give you starter code and validation steps to run on your own machine.