Giovanni Mazza

I’ve been the architect and the lead developer of Skye, a voice-to-voice agent used by thousands of pupils in the UK and the US. The agent replaced human tutors as the company's tutoring provider, allowing it to revolutionize the tuition market and position itself as a leader in the new AI-first tuition market. The project is backed by the Gates Foundation in collaboration with Stanford and Cornell University.

I designed and developed Brijio, enabling AI agents to use your native browser via Safari or Chrome extensions, accessing your live session with cookies, logins, and more. This eliminates CAPTCHA challenges and bot detection barriers.

My responsibilities included creating the server-side components: a WebSocket relay for managing browser connections and routing tool calls, and an MCP server offering functionalities like page reading, clicking, and form filling. The project also involved developing a shared protocol, health endpoints, structured logging, and an integration test suite.

Brijio was released as an npm package and a single Docker image. Technologies used included TypeScript, pnpm monorepo, per-request WebSocket connections, and default configurations for local development.

As a technical architect, I've been tasked to re-architect the old platform to improve scalability, reliability, and observability.

I chose AWS serverless technology to achieve our goal, and I implemented the BfF pattern with AWS API Gateway and the SAGA pattern using AWS Step Functions. The observability goal has been reached using Amazon CloudWatch to trigger alarms and notifications when an unhandled error happens.

An intervention program software that helps pupils boost their KS2, SATs, and GCSE performance with the help of real tutors across the word. The platform manages tutors' contracts, students' assessments, teachers' planning sessions, and admin operations.

A math resource hub for maths to help teachers to improve their lessons. The CMS organizes and correlates resources, and the API serves millions of requests using an efficient caching mechanism to reduce the load on the database.

An ERP system to help small/medium-sized company to manage their business. The software manages resources, accounting, warehouse, tasks, tickets, etc. As a sole developer, I built the framework and most of the modules for the ERP system.

iOS app to calculate the depth of field (DoF). The app scraps data from the web to build a database of cameras and lenses to help the user carefully calculate the DoF without knowing the spec of the camera or the lens. The app searches, for example, images taken with the same camera and the same lens via Flickr API.

A microservices ecosystem built to set up CMS, eCommerce, newsletter, and other services by activating just the services the customers need. The ecosystem was coordinated by a centralized house-made orchestrator.

The Subnotify app indexes the followed TV shows and gets a notification on iOS devices when a new subtitle is available. The app implemented a few smart features to match user preferences with TV shows, maintaining a database of TV shows, episodes, and actors to suggest new releases to users actively.

I designed and implemented an event-based evaluation suite for Skye, Third Space Learning’s AI voice tutor, used in real classroom sessions. The system validates AI tutor behavior against editable pedagogical rubrics, combining synthetic scenario tests with production trace evaluation.

The suite supports simulated tutoring scenarios, expected-response checks, and post-hoc scoring of real session logs across dimensions such as pedagogical appropriateness, responsiveness, safety, and rule adherence. Skye emits structured analytic events throughout each session, which are streamed through Fluent Bit into AWS Kinesis and processed by Lambda into the data warehouse. This allows the team to evaluate agent behavior at scale, monitor quality trends, and iterate quickly on tutoring rules as the AI system evolves.

My role covered the architecture, event model, evaluation strategy, AWS data pipeline integration, and the rubric-driven testing approach used to make AI tutor quality measurable and repeatable.