AI agent & narrative systems consulting
AI that respects the rules.
And remembers what matters.
We help teams building AI games, interactive simulations, and stateful agents turn unpredictable behavior into something they can inspect, improve, and test.
Discuss your projectStart with the behavior you need to fix.
When a good demo meets a difficult session.
- The AI changes the outcome.
- Separate what the model can describe from what rules and application state have already decided.
- A character knows too much.
- Define who can access a fact, how they learned it, and whether it is knowledge, a rumor, or a belief that needs correcting.
- Yesterday’s choices disappear.
- Connect memory to a durable record of people, events, and state changes that the next interaction can use.
- A prompt change breaks something else.
- Build repeatable scenarios that check behavior across turns and make regressions visible.
One failure.
A focused engagement.
A reliability audit and implementation sprint, scoped to your system.
We start with a behavior that matters to your users and agree on what a successful correction would look like. The work can cover architecture, state and memory design, tool permissions, or evaluation.
- Reproduce it. Turn the failure into a scenario we can run and inspect.
- Find the boundary. Decide which responsibilities belong in code, stored state, retrieval, or generation.
- Implement the correction. Address the agreed failure in your application.
- Leave the proof. Deliver regression tests, before-and-after results, and a handoff covering remaining limitations, latency, and cost.
Scope, timing, and fees are agreed before work begins.
Table for One is a solo RPG platform that connects an AI game master to Ironsworn rules, dice, persistent campaign state, and a journal. Building it means working through the same problems this service addresses.
Keeping the narration answerable to the game
If a roll resolves as a miss, fluent prose alone cannot make it a success. The architecture gives mechanics and state their own authority, and the test harness can exercise the real play pipeline with controlled dice outcomes.
Inspect a recorded turnResearch prototypes: knowledge boundaries and writer-authored tests
Separate prototypes explore who may know a fact, when a correction reaches a character, and how a writer’s narrative boundary can become an executable policy check. These are bounded demonstrations, not a production-ready standalone SDK.
Work directly
with the builder.
I’m John Demic, the builder of Table for One. My work on the project spans the AI GM harness, rules and state integration, narrative continuity experiments, and evaluation infrastructure.
You’ll work with me to define the problem, inspect the evidence, and decide what to change. Table for One is the case study behind this offer; the engagement is built around your product and its constraints.
What should your AI
be able to get right?
Tell me what you’re building, one behavior that is failing, and what you’ve already tried. Include your timeline if you have one.
Discuss your project hello@tableforone.ai