Skip to content
Shekhar Kumar

All projects

Work · Shipped

Fleet

GitHub and Booking.com rebuilt as environments for agent evaluation and RL.

Role
Development lead
Organisation
Pure Billion Technologies
Updated

TL;DR

  1. Led development of an agent evaluation and reinforcement-learning platform.
  2. Recreated production applications such as GitHub and Booking.com as interactive UI environments.
  3. Agents execute multi-step tasks, observe application state, and receive task-level feedback.

Scrub an agent's episode

Synthetic example

task Star the repository acme/widgets, then open its newest issue.

github.local/
github.localsearch…

Recent activity · Your repositories

observation
Signed-in home page. Search box in the header.
action
click(header.search)
reward
+0.0 · total 0.0
A synthetic episode on a GitHub-like environment. Each step pairs what the agent observed with what it did; the task-level reward arrives only when the environment confirms the task is complete.

The problem

To train and evaluate agents that operate real software, you need environments that behave like that software: real pages, real state changes, real multi-step tasks.

Live websites can't give you that safely or repeatably, so Fleet rebuilds the applications themselves as environments agents can act in and be scored on.

My role · Led development of the platform.

How it works

  1. Recreate the application

    Apps such as GitHub and Booking.com are rebuilt as interactive UI environments with their own state.

  2. Define a task

    Each task describes a multi-step goal an agent must accomplish inside the environment.

  3. Observe and act

    The agent observes the application state, takes an action in the UI, and observes again, step by step.

  4. Feedback from state

    The environment checks its own state to produce task-level feedback used for reinforcement learning and evaluation.

Key decisions

  • Rebuild the applications as environments

    instead of: Point agents at the live websites

    Owned environments can expose their state for scoring and be run again and again without touching real users or accounts.

  • Task-level feedback from application state

    instead of: Judging transcripts after the fact

    Whether the repository is starred or the booking exists is a fact the environment can check directly.

Outcomes

  • Environments recreating applications such as GitHub and Booking.com for agent training and evaluation.
  • Agents can execute multi-step tasks, observe state, and receive task-level feedback for reinforcement learning.

Stack

  • RL environments
  • Agent evaluation

Built at Pure Billion Technologies; client details abstracted.