Use Case

1. Prompt and Trajectory Creation

The goal is to collect human-generated plans and browser actions for everyday tasks-like ordering food or bookingflights-using a secure browser extension. This extension captures multi-tab activity, annotates data safely, and storeseverything in the client’s data warehouse.

2.Agent Plan and Execution Evaluation

2.A. Evaluating the plan generated by an AI browser agent

The objective is to evaluate the agent’s plan for each collected task, with every step annotated and checked for mistakes.

‍

2.B. Evaluating the execution of an AI browser agent

Objective is to assess execution of the plan generated by the agent. This is being done by the human experts.

‍

The Bottleneck: Data Quality in Intelligent Agent Training

Training your intelligent agents to navigate the digital world effectively hinges on the quality of their learning data. Imagine trying to teach someone to drive with blurry maps and incomplete instructions – the outcome wouldn't be pretty. Similarly, flawed or incomplete browser trajectory data leads to:

‍

The Deccan AI Difference

• Secure, Client-Side Data Collection and Annotation
  We developed a browser extension that captures trajectories—including screenshots and user actions—across multiple tabs,with all annotation performed client-side. This ensures data never leaves the client’s environment unless explicitly permitted.Data is stored directly in the client’s data warehouse, maintaining full control and compliance with internal policies.
• End-to-End Workflow Coverage
  Our tools enable the collection of both human and agent-generated plans, supporting a wide range of use cases (e.g., e-commerce, travel, support). Annotators can label each step, flag errors, and add context, ensuring rich, actionable datasets.
• Advanced Plan and Navigation Evaluation
  We implemented a dual-layer evaluation framework:

• Scalability and Flexibility
  Our architecture supports high-throughput data collection and annotation, with flexible integration into existing clientworkflows and data infrastructure.

‍

Our Gold Standard: End-to-End Quality Control (QC)

We don't rely on assumptions. Our multi-stage QC process ensures the highest level of annotation accuracy and consistency:

Prompt and Trajectory Annotation Workflow

‍

ModelEvaluationWork ow

‍

‍

‍

‍