IRIS — Natural-language data querying powered by AI

Overview

An enterprise AI platform that lets non-specialists query complex databases through ordinary conversation. A question like "which department had the highest attrition this year?" is translated into optimised SQL and answered with the analysis result.

Its ontology-driven architecture separates business concepts from the database schema, so the system can be deployed to a new industry within two weeks without changing the domain layer.

Core architecture

An eight-stage query pipeline

B-1 Orchestrator     →  request routing and context management
B-2 Intent parser    →  LLM-based classification of query intent
B-3 Ontology resolver →  extracts business concepts and maps them to tables
B-4 SQL generator    →  LLM-based generation of optimised SQL
B-5 SQL validator    →  security and safety checks (whitelist)
B-6 Query executor   →  execution on Supabase/PostgreSQL
B-7 Response formatter →  LLM-based natural language summary
B-8 Query logger     →  audit log plus self-learning data

The ontology engine

Business concepts — classes, properties, aliases and join paths — are defined in a YAML ontology. Query logs feed back into that ontology, giving the system a self-reinforcing structure.

# example ontology definition
classes:
  - name: Employee
    table: employees
    aliases: ["직원", "사원", "임직원"]
    properties:
      - name: department
        column: department_id
        join: departments.id

Key features

Analysis tiers

• Tier 1: free-form ad-hoc natural language queries
• Tier 1.5: structured comparative analysis
• Tier 2: trends, cohort analysis and rankings

Security and access control

• Five-level RBAC: data access scoped by role
• Department filtering: out-of-scope data blocked automatically
• PII masking: sensitive data de-identified automatically
• SQL whitelist: only approved patterns are allowed to run

Self-learning

• Unmatched queries collected and analysed automatically
• Golden query library (over 100 curated queries)
• Ontology extended automatically from usage patterns

Workflow

• Write-back support for three key fields
• Approval workflow integration
• Microsoft Teams integration (planned)

Tech stack

Next.js 16 + TypeScript 5 + React 19
Tailwind CSS 4 (UI)
Supabase PostgreSQL (database)
Claude Sonnet 4 / Opus 4.6 (LLM)
React Flow (pipeline visualisation)
Recharts (data charts)
js-yaml (ontology parsing)
Vercel (deployment)

Roadmap

| Phase | Scope | Status | |------|------|------| | Phase A–D | Core pipeline, HR ontology, RBAC | ✅ Done | | Phase E | Multi-turn conversation, RAG, streaming, automatic ontology growth | 📋 Designed | | Phase F | Multi-agent, IRIS Workshop (low-code), external integrations | 🔮 Planned | | Phase G | Multi-domain platform, domain marketplace | 🔮 Vision |

Design philosophy

IRIS follows an ontology-first architecture inspired by Palantir Gotham. Business concepts, not the database schema, drive SQL generation — which means a domain expert can extend the system without engineering cost.

The target for onboarding a new domain is four weeks end to end: analyse the database, write the YAML ontology, deploy.