Gabriel C.

ClassGenie.ca, a platform created to optimise language learning in school systems.

Some of the challenges we faced:

Creating custom voice agents with minimal lag and maximum quality. Things that seem obvious when you see the end product are not when you develop it. For example, I used some live models like Gemini 2.5 live, which output text and audio at the same time.

Well, funnily enough the text that Gemini 2.5 live outputted was not a 1:1 match with its audio, so I had to find a way around it by using a different transcription model (in this case, Elevenlabs scribe)

Quality control agents. It's essential to have safeguards to control the quality of agents. But infinite quality control makes the agents slow and unusable. A key challenge was to find the right amount of safeguards that guarantees the system working well, while being as transparent as possible to the end user.

Fallback models (I've learned this the hard way): Imagine your whole platform depending on a company like OpenAI not glitching. That's why almost all of our custom agents have fallback models.

Chesskool is a paid-membership web app I built using React and Meteor.js, where users learn chess through spaced repetition. Over 2,500 users.

Startup video streaming consultant:

Application development.

Product iteration across multiple phases to achieve a functional, incremental MVP.

Product–market fit testing, from new idea conception to market feedback through experiments (Facebook ads, Instagram ads, etc.).