Chapter 06AI-Native Runtime

AI Runtime — AI-native architecture

AI is a built-in runtime, not an add-on. Atan is an OS-level assistant, not a chatbot.

AI is a runtime, not a plugin

The AI capability center and the CLI share one capability definition — what AI can do, CLI can do

OS capability registry

Business capabilities register to the OS via a unified interface

AI Runtime dispatch

The LLM understands intent and dispatches registered capabilities

Atan execution

Natural language → capability call → structured response

Six AI capabilities

OCR · extraction · anomaly detection · forecasting · Q&A · decision support

OCR document recognition

Invoices, contracts, and reports structured automatically

PaddleOCR

Information extraction

Unstructured text → structured business data

LLM

Anomaly detection

Realtime energy and device anomaly alerts

Time-series models

Trend forecast

Energy and emissions trend analysis

ARIMA + LLM

Intelligent Q&A

Policy interpretation, data queries, report generation

LLM + RAG

Decision support

Charge/discharge strategies and efficiency recommendations

Optimizers + LLM

Atan · conversation demo

Not a chatbot — an OS-level assistant that operates the business system

AtanAI Assistant

How is this month’s emissions trend?

This month’s total emissions: 892.4 tCO₂e, down 12.3% YoY. Main abatement sources: 1. Solar substitution → -156 tCO₂e 2. Smart HVAC optimization → -89 tCO₂e 3. Lighting retrofit → -34 tCO₂e Watch: Building 3 HVAC is 18% high. A work-order suggestion has been created.

Find the latest carbon-trading policies

Found 3 recent carbon-trading policies: 1. 2026 National ETS Allowance Allocation Plan (2026-02-15) 2. Interim Measures for Carbon Trading — implementation rules (2026-01-20) 3. Voluntary Emission Reduction Trading — supplementary rules (2025-12-30) Which one should I explain in detail?

AI deployment architecture

Supports on-prem inference for data-security requirements

Ollama on-prem inference

LLMs run on local Mac Studio / GPU servers; data stays on the LAN

PaddleOCR private

OCR models deployed on-prem; document data stays secure

Graceful fallback

If AI is down, a rules engine takes over so operations continue