Simulated people to train your AI agents.
Real conversations between people and agents are scarce. Alyosha, Jean’s model of human behavior, plays the user: an environment for reinforcement learning, and a source of synthetic training data.
Trained on people, not prompted to act like one.
Assistant models are trained to be helpful. Ask one to play a frustrated user and it stays polite. Alyosha is trained on what people said and did, so your agent meets the users it will meet in production.
See which users break your agent.
Every episode has a goal the agent cannot see and a check on the outcome, so the reward is verifiable. Pass rate comes back split by the kind of user.
Episodes you can train on.
Every episode comes with the persona, the hidden goal, the full transcript and the reward.
- Rollouts for reinforcement learning
- Synthetic conversations for fine-tuning
- A held-out set for evals
How it works
- 1Your domain
Adapt
Alyosha is tuned on your logs and tasks, so it behaves like your users.
- 2At scale
Simulate
It plays the user in every episode, each with a persona and a hidden goal.
- 3Your loop
Train
Use the episodes as RL rollouts, fine-tuning data or a held-out eval.
- 4Against people
Check
A slice of the same tasks runs with real people, so you can trust the reward.
Why labs use it
- DiverseExpert, novice, vague, stubborn: a sweep, not a persona.
- Held fixedA frozen environment you can train against for weeks.
- Your domainAdapted to your logs, not the internet's.
- MeasuredOpen evals, same tests as everyone.