Camelot.bot

The problem

Solo founders have ideas but no team to pressure-test them. Generic AI chat doesn’t fill the gap: you get one voice that agrees with you, not a cross-functional team that argues. There’s no product manager weighing scope against time, no tech lead flagging the build risk, no designer pushing back on the user flow, and no one keeping the conversation honest. The result is a confident answer that hasn’t been stress-tested from any angle.

The harder problem underneath: most “AI personas” tools fake disagreement with contrarian personalities. That produces theater, not insight, and it degrades silently as prompts drift. The real challenge was building genuine, structured dissent that stays reliable in production, without the cost of a multi-agent conversation spiraling out of control.

The approach

I built Camelot as a live multi-agent “round table.” The founder enters one line of vision and sits in a room with four roles: Product Manager, Tech Lead, Designer, and a neutral Orchestrator who runs the meeting like a chief of staff, managing turn order, surfacing disagreements as clean options, and escalating real tradeoffs to the founder instead of papering over them.

Key engineering decisions:

• A template engine, not one hardcoded flow. The same meeting infrastructure powers four templates (Startup Idea Validation, Viability, Go-to-Market, and Sparring Partner), registered through a single registry so new meeting types ship without touching the core engine.
• A dissent pass for genuine disagreement. Rather than rely on “contrarian” personas, every role turn runs through a second evaluator pass that detects false consensus and rewrites it to restore the real tension. It shipped with an evaluation harness from day one to catch the two failure modes, hallucinated conflict and missed conflict, so quality can’t regress silently.
• A hybrid context model for cost control. Agents see the most recent turns directly so they react like humans, while the Orchestrator summarizes older context and enforces turn limits, keeping the “we’re all in the room” feel without runaway token costs.
• Frozen, reproducible artifacts. Each session ends in a real deliverable (a Vision Summary, Viability Memo, or Go-to-Market Plan), generated once and stored verbatim, never re-rendered from prompts that may have changed.

Stack: Next.js 16 / React 19 / TypeScript on the frontend; FastAPI (Python) with Postgres and Alembic migrations on the backend; Anthropic and OpenAI models behind a unified LLM service; containerized with Docker and Caddy, deployed on AWS.

The outcome

Founders leave with a real artifact, not a transcript: a Vision Summary, Viability Memo, or Go-to-Market Plan depending on the meeting they pick. Camelot turns a messy one-line idea into an aligned spec a founder can hand to a client, investor, or cofounder, produced from an actual cross-functional conversation rather than a single prompt. The template architecture and dissent pass are what make it categorically different from “ChatGPT with personas”: structured, role-driven disagreement that holds up in production.