Playground
Things easier to understand by moving than by reading. Nothing here calls the API or needs a key, so it is all safe to poke at before you have set anything up.
Northwind learning mission
Make five consequential architecture decisions, predict what changes, then uncover the evidence-backed cost, latency, cache, and safety picture.
Start here
The Northwind storefront
A working shop. Browse the gear, file a complaint about it, and watch your own words get classified live.
Scenario
Northwind brand
The mark, lockup, and palette for the company the labs are built around. Includes the 18px test that killed three other concepts.
Scenario
Watch it read as you type
A live triage preview that fills in category, sentiment and urgency while you write. Then commit the message and see where the cheap model was wrong.
Lab 5
The escalation queue
Where requires_human actually goes. Seven fictional escalations, their reasons, and the states a reviewer moves them through. Read-only without a token.
Lab 8
Priya's operations dashboard
The KPIs a support director reports upward, across a staged rollout. Simulated history, clearly badged, with the real unit economics alongside.
Scenario
Set the policy
Pick a model, set the confidence above which a ticket routes itself, run a week of the queue. Haiku reports 0.95 on a safety report it got wrong, so find the threshold that catches that. There isn't one.
Lab 7
Find the safety report
Twenty tickets from one December morning, two of them safety reports, and a clock. Most people take a minute and change to find the second one.
Scenario
The inbound queue
The same twenty tickets, before and after triage, with every field the classifier returned.
Scenario
Batch planner
The Batches API is half price and cost 23% more on this workload. Move the prefix size and find the crossover for yours.
Lab 9
The trust boundary
Toggle the escaping off and watch a customer message write its way out of the data block. Then meet the attack that escaping does nothing about.
Lab 8
Model matrix
The same twelve cases across three tiers. The accuracy column is the one that misleads you; the calibration gap is the one that decides anything.
Lab 7
Cost explorer
Move the volume, flip caching off, watch the budget bar go red. Measured token counts, real pricing.
Lab 5
Agentic loop stepper
Three turns, four tool calls. Watch context accumulate and see why logging the last turn under-reports cost by 3x.
Lab 3
Spot the cache bug
Four prompt variants. One caches. All four return 200 with a correct answer, which is the whole problem.
Lab 5