Heart‑Eco · HEART AI — Agentic Economic Analytics

The Problem

Professor Khurshid Ahmad's research team needed a private, secure AI assistant that could interrogate decades of environmental sustainability data using the HEART Score Economic Model. The key challenge: the AI had to strictly use provided datasets, refuse external knowledge, and handle year-range validation — no hallucinations allowed.

Existing tools couldn't guarantee data grounding, and commercial analytics platforms lacked the specificity needed for HEART metrics.

Our Solution

Two-Model Architecture

Model 1 on OpenAI Agent Builder for descriptive / perspective / diagnostic answers; Model 2 on n8n + GPT for predictive forecasts grounded in ensemble ML. This separation ensures each model operates within its domain of expertise.

Strict Tool Routing

Numbers come from numeric master sheets only; methodology from doctrine docs; narrative from forecast commentary — preventing the model from improvising facts. Every data point is traceable to its source.

Forecast Transparency Rules

Years 2026–2030 are auto-labelled as forecasts with explicit fallback behaviour when narrative is missing. Users always know whether they're looking at historical data or projections.

Dashboard Deployment

Next.js 14 + TypeScript + Tailwind + Recharts dashboard, deployed on Vercel as part of the HeartEco product. Interactive visualisations let researchers explore data without writing queries.

Domain-Faithful Boundary

Restricted knowledge boundary in the system prompt — refuses unrelated queries and never uses web knowledge. The AI stays in its lane and tells users when a question falls outside its domain.

Tech Stack

• AI: OpenAI Agent Builder, n8n workflows, custom RAG pipeline
• Frontend: Next.js 14, React, TypeScript, Tailwind CSS, Recharts
• Data: Vector database, pandas, custom ingestion scripts
• Deployment: Vercel

---

Need an AI system grounded in your data? Talk to us about RAG.