Ex-CEO, 12 YOE, but "unhirable". So I built a complex D&D Virtual Tabletop with NLU & Realtime in a weekend using AI. Here is the stack.

I have 12 years in IT. I’ve managed teams of 50+, handled $3M budgets, and held the "CEO" title for the last 3 years. Paradoxically, that title killed my employability. Recruiters see "CEO" and think "flight risk" or "forgot how to code". So I stopped applying. I decided to prove that a Senior Architect + AI Agents can outperform a traditional dev team.

The Build: A "Roll20 Killer" Backend I didn't build a simple CRUD wrapper. I built a complex TTRPG platform (think D&D/Pathfinder) with heavy business logic.

The Tech Stack (The "Meat"):

Backend: Python (FastAPI) — chosen for speed and async capabilities.

Realtime: WebSockets/SSE via Redis Pub/Sub for syncing dice rolls and game state across clients.

Database: PostgreSQL with complex JSONB schemas for Skills, Effects, and Spells to support different game systems (d20, d100) dynamically.

The "Impossible" Stuff I Built Fast with AI: Usually, these features take weeks of fine-tuning. With Cursor/Windsurf, I shipped them in hours:

The "Roll Engine" & Math: I implemented a complex dice roller that handles Advantage/Disadvantage, crits, and follow-up rolls (like damage concentration checks).

AI Workflow: I fed the SRD rules to the agent, and it generated the recursive calculation logic + 100% unit test coverage instantly.

Lore-Aware NPCs with "Fog of War" (Scoped RAG): The GM uploads their world lore (PDFs/Text), and NPCs understand who they are and the history, but without being omniscient.

The Challenge: A peasant in a tavern shouldn't know the King's secret war plans just because it's in the vector database.

My Solution: I architected a scoped knowledge retrieval system. The AI only injects context relevant to the NPC's location, rank, and the current scene. The NPC stays in character and doesn't spoil the plot.

NLU (Natural Language Understanding): The system parses text like "Cast Fireball on Orc" into structured JSON actions {skill_id, targets[], rank}.

AI Workflow: I used LLMs to generate the training dataset and the parsing logic, connecting it to the game mechanics seamlessly.

Billing & Multi-Tenancy: Built a credit-based billing system (Token Bucket rate limits) and organization-level multi-tenancy from scratch to manage LLM costs per campaign.

My "Vibe Coding" Workflow I act as the Principal Architect.

I write the Contract/Spec (like the ones attached to this post).

I feed it to the AI Agent.

AI scaffolds the boilerplate, DB models, and basic logic.

I review, refactor, and handle the complex edge cases (like race conditions in the billing ledger).

Why I am sharing this? I’m launching this as a service — BuildItFast.ai. I realized that founders don't need "code". They need complex MVPs shipped yesterday.

Question for the community: For those selling high-end dev services: do you show clients the code complexity (like my Scoped RAG/Realtime architecture) or just the result? I feel like non-tech founders get scared of the "how", but tech founders love it.

Roast my stack or my approach.