Utkarsh Bali
Utkarsh Bali
Work

Agent QAIndependentIndependent build2025Shipped

LLM Multi-Agent QA System

Give it a plain-English task on an Android phone, like find the best pizza place near me and start Google Maps to it, and it does it. Four agents plan, execute, verify and supervise the run.

>99% deterministic execution, internal benchmark

Problem

Most agent tests fail in boring ways: a bad click, a missing state check, or a recovery path that was never designed.

Approach

I built this for Android flows.

Splitting the work across four agents mattered more than making any single one smarter.

How it works

An agentic ADB pipeline that turns a spoken goal into Android actions, with four agents coordinating over a message bus.

Four agents in lockstep, sixteen steps, no diagonals

  1. 01

    Planner agent decomposes natural-language test goals into executable UI steps, replanning when the app diverges.

  2. 02

    Executor drives Android Debug Bridge with GPT-4o and OpenAI Vision reading the screen, while the verifier checks state transitions.

  3. 03

    Supervisor logs decisions and owns retries, failure recovery, and deterministic completion criteria.

GPT-4o, OpenAI Vision, ADB
Architecture diagram of the multi-agent QA system, showing the planner, executor, verifier, and supervisor agents communicating over a message bus within an episode loop

Impact

The same instruction produces the same run, which is what makes a flaky UI testable at all.

Built with

Python · GPT-4o · OpenAI Vision · ADB · UI Automator · Android SDK