What I Built

AI Boardroom OS is a multi-agent business decision simulator that runs on a local LLM. You describe a business decision, optionally upload supporting documents, and a virtual boardroom of 10+ executive personas (CFO, CTO, CMO, and others) analyzes it, debates it, and hands it to a CEO agent that synthesizes a final recommendation. A last validation stage checks that the decision is coherent.

Who would pay: solo founders, small agencies and consultancies who want a structured second opinion on pricing, hiring, or go-to-market calls but can't afford an advisory board or send confidential numbers to a third-party API. Everything runs on the user's own machine.

Side hustle angle: [your model, e.g. "sell it as a one-off 'decision review' service at ₹X per session", or "offer a self-hosted license to small consultancies". Fill in what you'd actually do.]

Code

How I Built It

• Model: Qwen2.5-3B-Instruct (open weights), 4-bit NF4 quantized, running on an RTX 3050 with 4 GB VRAM. I started with TinyLlama-1.1B, moved to Qwen2.5-1.5B, then to 3B as I got better at working around small-model limits.
• Pipeline: four stages: executive analysis → debate → CEO synthesis → decision validation.
• Small-model tricks: JSON-schema-constrained outputs, role-specific grounding hints, and compressed "cards" passed between stages so context stays small. A truncation-retry function handles incomplete generations.
• Retrieval: Sentence Transformers + ChromaDB for RAG over uploaded documents.
• Stack: Python, PyTorch, Hugging Face transformers, FastAPI backend (/boardroom/analyze), Streamlit multipage frontend with a landing page and a "Command Center" page.
• Modes: 2 intelligence modes, 3 response formats, and auto or manual agent selection.

Why Does Open Innovation Matter?

A closed API would have meant per-token costs on every multi-agent debate, which is a lot of calls for one decision, and my clients' business data leaving their machine. Open weights let me quantize the model to fit a 4 GB laptop GPU, constrain its output with schemas, and tune prompts per role. That is what makes a private, near-zero marginal cost advisory product possible for the small businesses I'd sell it to.