← WorksLinyun Jiang

AI Agent: Khora

Khora

AI Assistant

AI/LLM + Architecture

An LLM design system that generates buildable timber assemblies under a hard constraint on real stock. Every member must be drawn from three libraries totalling 18,401 real timber members, and no member may be reused within a design.

The system is a neuro-symbolic architecture: the LLM handles open-ended generation, deterministic code guarantees closure. Each design makes only two LLM calls; deterministic stages then carry out specification rebuild, real-stock allocation and structural repair. At the core of the system is a voxel intermediate representation. The outer loop optimises the deterministic repair rules, not the model weights. Acceptance is decided by a deterministic gate and three independent verifiers (voxel fidelity, physical compliance, material authenticity), not by model self-assessment.

Role

  • Designed and built the training pipeline and its scoring routine
  • Chain architecture and server
  • Interface and interaction design, front-end
  • The exhibition site
Instructor
Marios Tsiliakos, Stefan Bassing
Type
Academic
Date
2026.05 - 2026.10
Award
Computational Prize 2025-26, UCL The Bartlett School Of Architecture
Team Members
Linyun Jiang

Livekhora-ai.com, liveOpen in new tab ↗

Training architecture

Khora Training Process

UI / UX: the exhibition site

The site made for the exhibition at khora-ai.com: a poster with a QR code, and a page each for the training record and the models of the second and later rounds.

UI / UX: Khora AI

The assistant itself: a chat with the chain behind it, and a record of the briefs it was given.

Khora Advantage

The same brief, given to Khora and to a general-purpose model. What is measured is not taste but whether the design comes back buildable: every member drawn from timber that actually exists in the stock, no member reused, nothing left floating.

Next projectMy BabelAI/LLM + Education + Architecture, 2022My Babel, cover