The Company
Twenty Billion Neurons GmbH was an AI research startup founded in 2015 with the mission of making virtual fitness coaching accessible to everyone. The company specialized in artificial intelligence research and enterprise-scale software engineering, with offices in Berlin (R&D, Product) and Toronto (Research, Animation), employing around 40 people.
TwentyBN was funded by M12 (Microsoft's venture capital subsidiary) and advised by Yoshua Bengio, the 2018 Turing Prize recipient. The company was recognized as a CB Insights 2019 Top 100 AI Startup and a Gartner Cool Vendor in AI Core Technologies. Their clients included BMW, Microsoft, Google, Apple, IBM, Baidu, and Alibaba.
The company authored the foundational "Something-Something" and "Jester" video understanding datasets and was ultimately acquired by Qualcomm in December 2020.
The Product — Fitness Ally
The flagship consumer product was Fitness Ally, a mobile app that used machine learning for real-time video recognition to provide live feedback during workouts. The app was guided by Millie, a virtual being with personality and intelligence — animated using the Unity game engine — who simulated the experience of a personal fitness trainer.
TwentyBN built their datasets through a crowd-sourcing platform where people were paid to record themselves performing specific movements. This data was used to train neural networks capable of understanding and interacting naturally with humans in real-time. The tech stack consisted of a backend in C, Go, and Python, with the frontend built in Unity, all running on high-performance Linux workstations with dedicated GPUs.
My Role
I joined as an AI Engineer Intern, supervised by Florian Letsch and Cornelius Bohm, both Senior AI Engineers. I was a full member of the agile software development team in the R&D department, working on the Fitness Ally application using Python and JavaScript under Scrum methodology with 2-week sprints.
I was responsible for three main missions:
1. Challenge Mode
Implemented a new in-game feature called Challenge Mode that pushed users beyond their physical limits while receiving live AI-driven feedback from Millie. This involved integrating with the real-time video recognition pipeline to dynamically adjust difficulty and coaching based on the user's detected movements.
2. Voice Lines & Reactions
Integrated a wider range of voice actor recordings for Millie's dynamic interactions with users. These voice lines and reactions were recorded by professional voice actors and had to be programmatically mapped to specific user actions and states within the application's interaction system.
3. Visual Debugging & Flow Tool
Built a web-based visualization tool (in JavaScript) that allowed non-technical team members — product owners, VPs — to easily trace and understand user flow throughout the application. This tool was presented at sprint reviews and helped the team inspect the behavior of the software without needing to read code.
Results
The internship report received a score of 19.5 out of 21 from PUC, rated "Destacado" (Outstanding). The company evaluation was a perfect 12 out of 12, with top marks across technical competence, teamwork, communication, and professional behavior.
My supervisor wrote: "As an intern, Diego was a full member of our agile software development team. He contributed on a technical level, but was also a curious and proactive member who took part in improving our internal processes and communication structures. In team discussions, Diego was an active participant and shared constructive ideas with the team. We have perceived him as a very motivated person who strives to always improve his work and his way of working. We enjoyed having Diego with us."