UNDERSTAND THE MACHINE
How models, reasoning patterns, tools and environments come together in a run.
RLGym
A workspace for humans to teach AI agents how to reason, use tools, and recover when they get it wrong.
- ROLE
- Product Designer
- SCOPE
- Product thinking · UX · UI
- PLATFORM
- Enterprise AI
- TEAM
- Product · Engineering · AI
/projects/rlgym/hero.pngCONTEXT / 01
Think of it as a practice world for AI.
Instead of letting an AI agent learn inside a real enterprise system, RL Gym gives it a controlled environment where it can observe, act, receive feedback and try again.
DISCOVERY / 03
Two weeks.
No existing product.
Just the system to understand.
RL Gym was a 0→1 product. I started with the product specification, working sessions with the AI/ML and product leads, and one clear user group — annotators.
The first job wasn’t designing screens. It was understanding how the system behaved and where a human needed control.
UNDERSTAND THE ANNOTATOR
Where the human needs visibility, judgement and precise control.
TURN COMPLEXITY INTO A WORKFLOW
Convert many system states into a sequence an annotator could actually follow.
THE PROBLEM / 02
Training an agent is one problem.Understanding why it failed is another.
THE USER / 03
The human in the loop.
The annotator guides the agent from setup to usable feedback.
CONFIGURE
Choose the model, reasoning pattern, tools and environment.
MODELM1.1
REASONINGReact
TOOLS5 selected
ENVIRONMENTGoogle Suite
OBSERVE
Follow the trace: thought, action and observation.
THOUGHTIdentify user intent
ACTIONEmail_email
OBSERVATION3 emails returned
CORRECT
Edit, approve or evaluate the exact failing step.
FAILING STEPAction
ISSUEWrong tool selected
CORRECTIONTicketAPI
EDIT APPROVE EVALUATETRAIN
Turn the reviewed run into reusable feedback.
TRACEApproved
FEEDBACKStored
DATASETRL Gym Feedback
The first problem wasn’t visual. It was structural.
I explored different ways to organise setup, tools, runs and intervention before converging on a simpler flow.
/projects/rlgym/wireframes-setup.png
/projects/rlgym/wireframes-chat.pngThe system was complex.
The experiencedidn’t have to be.
SETUP—RUN—CORRECT
WHAT I DESIGNED / 04
Three moments
mattered most.
Set up the environment.
Understand the agent.
Intervene when it goes wrong.
SETUP
Set the model, reasoning pattern, tools and environment.
01 / 03
SETUP
Set the model, reasoning pattern, tools and environment.
/projects/rlgym/project-setup.png02 / 03
OBSERVE
Follow the trace: thought, action and observation.
/projects/rlgym/chat.png03 / 03
CORRECT
Intervene at the exact moment behaviour goes wrong.
/projects/rlgym/edited.pngFAILURE / 05
Then it picked
the wrong tool.
The run looked normal until one action sent the reasoning in the wrong direction.
1The run chose the wrong tool for a flight booking request.
2The annotator opens the action and inspects the decision.
3The right tool is selected before the run continues.
4The corrected result is ready for approval.
/projects/rlgym/wrong-tool-call.png
/projects/rlgym/annotator-clicks-edit.png
/projects/rlgym/select-right-tool.png
/projects/rlgym/final-result-approved.pngEVALUATE / 06
A correction still
needed a verdict.
Once the run was fixed, the annotator could evaluate the result and close the session with structured feedback attached.
/projects/rlgym/evaluation-hero.pngLEARNINGS / 07
Designing AI meantdesigning uncertainty.
The hardest part wasn’t making the system look intelligent.
It was making its behaviour understandable enough for a human to trust, inspect and correct.
TRACEABILITY
Make reasoning visible enough to inspect.
CONTROL
Let people intervene at the exact step without restarting the run.
REPEATABILITY
Turn corrections and evaluations into feedback that can be reused.
Complex underneath.
Clear where it mattered.