Official GitHub In-Depth Architecture
System One Python Adapter (LLM Fallback) →
Open-source drop-in replacement for TypeSafeClient backed by standard LLM APIs, enabling offline development and staging environment simulation.
Verified Demos & Architecture Patterns
To maintain strict developer credibility, this directory indexes only empirically verified demonstrations and design patterns backed by official repositories and TypeSafe AI documentation.
Official Reference Demo Source: TypeSafe AI
Smart Home Assistant (Speculative Fan-out Reference)
The official Smart Home Assistant demonstrates how a voice or text command like "Turn off all lights and set the bedroom AC to 22°C" is processed without multi-turn LLM agent loops. The system fans out dozens of typed questions in parallel—evaluating lighting, climate, security, and audio states concurrently.
Instant Concurrency: Single API call answers 15+ sub-device intent questions simultaneously.
Deterministic Parsing: Directly outputs typed Choice and Score objects with zero token hallucination.
Low Latency: Sub-150ms roundtrip execution suitable for real-time edge devices.
Core System One Design Patterns
Pattern 01
Speculative Intent Fan-out
When a user input arrives, the application immediately dispatches multiple parallel typed queries across multiple domain questions (e.g. intent classification, entity extraction, sentiment thresholding). Jev returns calibrated scores simultaneously.
const [intent, urgency] = await Promise.all([...]) Pattern 02
Guardrail Scoring & Routing
Use high-speed Score primitives (0.0 to 1.0) to gate risky operations or route requests between low-latency local paths and expensive long-thinking models.
if (safetyScore > 0.95) executeFastPath() Pattern 03
Drop-in OpenAI Adapter Pipeline
Wrap existing generative LLM chains with the official Jev adapter to replace text-generation intent parsing with structured decision evaluations without rewriting application logic.
import { createJevAdapter } from '@typesafe-ai/adapter'