Writing about AI, engineering, and building things.
What three people shipped between April and September: a public launch with paid plans, characters with minds, undo for an entire world, a tabletop mode, visual novels, and one week where our own retry logic tried to take down the database.
I tried to make the host of our Discord run on my own GPU. Most of what went wrong wasn't the model. It was my data, my tooling, and a loss curve that said 'done' while every step was NaN.
RoleCall's VN mode turns streaming roleplay into sprites, backgrounds, choices and endings. The hard part was never the renderer. It was deciding what happens when the model, the author and the player disagree.
Every RP benchmark is either vibes or tests the wrong thing. I built one that tried to do better, and the real lesson was about how badly LLM-as-judge fails on subjective tasks — until 162 community voters fixed it.
Inside RoleCall's DM system — a story director that plants seeds, runs a shadow world, and never spoils the narrator.
A sidecar AI that reads, writes, and organizes lorebooks in real-time — so the narrator never forgets a name, a place, or a promise.
RoleCall isn't a chatbot with a system prompt. It's an agentic system where multiple AI models coordinate per turn — planning, retrieving, generating, and updating game state.
How RoleCall routes inference across internal servers, a load-balanced proxy, and user-provided API keys — with E2E encryption, tier enforcement, and graceful fallback.
How RoleCall's 27 reactive trackers turn narrative text into combat encounters, relationship meters, inventory, and a living world — all driven by AI directives.
Static system prompts are dead. Dynamic world-building with structured lore is how you make AI characters feel alive.