True Bots
The problem. Game bots are scripted. Players learn the script, and the bot stops being interesting. We set out to replace scripted bots with ML agents that play at human level and can be trained piece by piece.
What I built. We were a core team of three. I was co-founder and COO: fundraising, investor relations, decks, projections, studio partnerships, and the company's admin, taxes and filings.
I also wrote the tools the work ran on. Training Studio was the desktop control panel our ML engineers trained on, 21,000 lines that I hand-rolled piece by piece:
- a live, interactive interface for the YAML training configs;
- a script that extracted the validation rules from ML-Agents itself, so they followed new versions instead of drifting;
- a validator that checked the training environment against them;
- an environment builder that built the right environment when one wasn't there.
Training Studio covered real-time agents. For turn-based games I built the LLM Bridge: a Python tool and a C# Unity plugin that open a channel to a local or frontier model and relay game state and moves back and forth.
What happened. We shipped six public demos and signed two studio partnerships. We were selected for La Guilde du jeu vidéo du Québec's Parcours Innovation cohort (CAD $4,000 and an OVHcloud infrastructure partnership) and for Founder University Cohort 10. The company dissolved in February 2026; my co-founder and I are still on the project.
What I learned. Build for the automation you can see coming. I gave Training Studio command-line and API entry points from the start, because we expected LLMs and agents to drive training before that was a common thing.