AngelQ: An AI-powered browser for kids
Vention provided three months of engineering support to AngelQ, an AI-powered browser for kids, through the In:Vention Incubator program.
The team automated CI/CD pipelines, reduced audio file sizes by 5-10x, improved LLM token streaming reliability, and built user management for AngelQ's first B2B school partnership.
Client overview
Key stats
Partnership duration
3 months
Team headcount
5 engineers
Areas covered
Backend, AI, mobile, DevOps
Manual CI/CD pipelines
Fully automated with GitHub Actions
The challenge
Our solution
Adding user management for the B2B direction
During its partnership with Vention, AngelQ entered the B2B market and struck a deal to supply its AI-powered browser to an American school. The move introduced new platform requirements, including structured user management and expanded notification workflows.
To support school environments, Vention implemented additional functionality alongside the browser’s code refactoring.
Vention:
- Designed and implemented authentication logic for business users, including admin, teacher, and student accounts.
- Redesigned the notification system so that not only parents but also relevant school staff are included in the reporting loop when unsafe information is detected.
New capabilities equipped AngelQ with the functionality needed to support both B2B and B2C usage scenarios.
Upgrading AI with better streaming and test reporting
Before Vention joined the project, the browser sometimes cut off token streaming during response generation or crashed when an empty prompt occurred (for example, when a voice prompt button was pressed by accident).
Vention reworked the token streaming flow (which controls how the AI generates and delivers responses) to handle both token input and output more reliably. The updated logic prevents interruptions during response generation and improves the stability and accuracy of the browser’s responses.
Vention’s AI engineers also improved testing and reporting for the LLM pipeline. Prompt coverage was difficult to evaluate, and test results were spread across different runs.
To address this, the team introduced several improvements:
- Augmented AngelQ’s CLI with promptfoo, an open-source LLM evaluation and testing framework, allowing the team to test how the LLM responds to different prompt inputs.
- Implemented a reporting system that consolidates all test runs into a single report containing input text, output text, and test results.
Better prompt coverage and centralized reporting now make it easier for the team to evaluate model behavior and detect issues earlier in development.
Upgrading the iOS/iPadOS app
Vention implemented several improvements for the existing native mobile apps:
- Bridged a Rust-based XC framework for native iOS/iPadOS development via UniFFI to leverage Rust’s memory safety and reliability for a children-oriented product.
- Added video player errors to the app’s analytics and error reporting.
- Created a framework for highlighting tooltips in mobile apps.
Beyond these improvements, Vention’s mobile engineer also worked to enhance the app's AI functionality, particularly text-to-speech.
PCM audio uses fixed packet sizes, while MP3 packets can vary. Because of this difference, handling MP3 streams required additional logic to detect packet boundaries and process incoming audio data correctly.
To address this challenge, the mobile implementation detects frame boundaries in MP3 files, caches incoming packets, and waits for the full audio stream to be ready before playback.
As a result, the mobile app can correctly process MP3 streams and stitch frames together during playback, reducing audio file size by 5-10 times while maintaining a seamless listening experience.
DevOps improvements
Before Vention got involved, most processes were manual and unoptimized. The team had to rely on local scripts for deployments, release projects as a single unit, and rebuild the entire system from scratch each time.
To address these limitations, Vention’s DevOps rolled out the following changes:
- Set up GitHub Actions for automated deployments.
- Redesigned the deployment pipeline so the system could reuse components that were not affected in the new version of the app, instead of rebuilding the entire solution from scratch.
- Added a separate pipeline that allows specific components to be deployed independently when needed.
The updated pipeline reduced manual work, improved deployment efficiency, and made the release process more reliable.
Results
Tech stack