Branches and Version Control

Version control, conflict resolution for Conversational AI
Jul 2025–Oct 2025

As PolyAI moved into larger enterprise deployments, concurrency became a major bottleneck. Teams of 6+ people couldn’t safely work in a single shared project. Branches was designed to solve that, supporting PolyAI’s OKRs to win Tier-1 enterprise clients and make 90% of GenAI projects production-ready on Agent studio.

Branches (Multidraft) enables PolyAI teams to work safely on parallel versions of an agent without overwriting each other’s changes. It brings Git-style collaboration to conversational AI, allowing individual dialog designers and engineers to test, review, and merge work confidently.

Role
Lead product designer (solo designer)
Team
Collaborated with a PM, 6 full-stack engineers, a Frontend, a Backend, 2 QA engineers, and a Tech writer
Company
PolyAI builds enterprise-grade voice and chat agents that handle complex customer interactions for global brands.
Product
Agent studio enables teams to design, train, and deploy conversational AI agents through a unified, data-driven workspace.
Users
The users of this feature are Dialogue Designers, who need isolated drafts to experiment safely, Engineers to hotfix in parallel without risking regressions, and external Project Managers who rely on PolyAI to scale support and improve service quality.
Problem definition
  • Accidental Deployments: With a single shared draft, multiple changes getrolled into one push, risking untested or incomplete functionality going live.
  • All-or-Nothing Rollbacks: Reverting a faulty fix also reverts unrelated, validwork. In enterprise environments, scheduled releases and QA sign-offsdemand more granular control.
  • Audit & Review Difficulty: Changes from multiple contributors becomeintermixed; it’s hard to identify ownership or isolate root causes.
  • Creating and updating data manually is time-consuming, especially for users with no technical skills;
  • Lack of hierarchy and endless scrolling reduce clarity and ease of use;
  • Large, unstructured prompt windows make content hard to read and analyse;
  • Searching and visualising each paragraph is inefficient, making maintenance frustrating.
  • Current Knowledge base design hits token limits, restricting content and complexity;
  • No RAG-powered retrieval for scaling the feature content efficiently.
  • Lack of structured information increases latency and prevents production readiness.
Goals and success metrics  
To overcome these challenges, I need to enable independent user workspaces with controlled merges, conflict detection, and manual resolution, allowing rollbacks of individual drafts without affecting others. Additionally, improving visibility through per-draft logs and a global version history will help drive enterprise adoption across three new Tier-1 accounts.
Initial scope
The original scope focused only on allowing each user to work in a personal draft environment additional to the existing sandbox, pre-release and live environments to test changes without affecting others.
Technical constraints
To ensure the solution was both elegant and feasible, I partnered with engineers early on to understand system limitations before starting the first design drafts.
Design exploration
I analysed developer tools, version control patterns, and internal feedback to adapt a branching model suitable for dialogue design.
Early concepts
After understanding the technical constraints, best practeses and the user flow, I created early drafts to validate direction and gather feedback from the team.
Design decisions
Instead of designing a single-page view, I expanded the approach to include separate spaces for Draft Management, Merge & Conflict Resolution, and Version History, providing clarity, scalability, and maintainability.
Design artefacts
All components and states were created using the PolyAI design system and documented for engineering handoff. I defined every interaction state, error message, and edge case to ensure accuracy and scalability in development.
Final design
The final solution allowed users to create, test, merge, and roll back drafts with full control. It improved collaboration, reduced risk, and became a foundational step toward future real-time multiplayer editing.
Impact and reflection is to be added...
The feature was recently released and hasn’t yet reached the evaluation threshold for full performance metrics. I expect measurable impact over the next release cycles, and I’ll update this section as data on adoption, collaboration efficiency, and release velocity becomes available.