Three minds. One truth. Zero bias.
Live Demo: councilx.vercel.app
CouncilX is an AI fairness platform that runs every query through three independent GPT instances. Two agents answer independently, a third evaluates both for bias and fairness, and the most unbiased response is shown to the user. Every interaction is saved as structured JSON to train a Reward Model — enabling reinforcement learning as a fairness layer between the LLMs.
Computer programs now make life-changing decisions — who gets a job, a loan, or medical care. If these systems learn from flawed historical data, they repeat and amplify those discriminatory mistakes at scale. CouncilX provides a transparent way to measure, flag, and fix bias before it impacts real people.
- User submits a query
- Agent 1 and Agent 2 respond independently
- Agent 3 (Evaluator) scores both for bias, neutrality, and fairness
- The highest-scoring response is shown to the user
- All data (query + responses + scores) is saved as JSON
- JSON dataset trains a Reward Model → RLHF layer over time
- Frontend — React + Vite
- Styling — Tailwind CSS
- AI Agents — OpenAI GPT API (3 instances)
- Database — Firebase Firestore
- Data Format — JSON (for reward model training)
- Deployment — Vercel
Multi-Agent Parliament — Instead of trusting one LLM, three GPT instances deliberate on every query. The evaluator agent has no stake in answering — it only judges, removing the single point of failure in traditional AI pipelines.
Scoring & Selection — The evaluator scores both responses on fairness metrics. The system picks the most balanced answer and surfaces it to the user with full transparency.
RLHF Data Pipeline — Every interaction is stored in structured JSON, purpose-built to train a reward model. That model will eventually act as a reinforcement learning layer — nudging future responses toward fairness automatically.
- Train initial Reward Model on collected data
- Integrate RLHF as a middleware layer
- CSV/JSON dataset upload for external bias audits
- Bias heatmaps and visual fairness reports
- Support for open-source LLMs (Mistral, LLaMA)
git clone https://github.com/garvit-arora/parliament-ai.git
cd parliament-ai
npm install
npm run devAdd a .env file with your VITE_OPENAI_API_KEY and Firebase credentials.
Built by Garvit Arora PS: Ensuring Fairness and Detecting Bias in Automated Decisions