Lumière AI

Most AI tools are chat boxes bolted onto workflows designed before LLMs existed.

Studied how power users chain ChatGPT, Notion, and a dozen scripts. The pattern was always: assemble context, run a step, save the output somewhere else. Nothing composed.

1. 01Agent runtime with tool-use and memory primitives\nVector store for retrieval over documents\nEdge functions for low-latency streaming\nProvider-agnostic model routing (OpenAI, Claude, Gemini)

Wrote the agent runtime first. Every UI surface — chat, docs, workflows — became a thin client over the same runtime, so features composed instead of forking.

Cost control. Naïve retrieval blew up token spend. Built a two-stage retriever that pre-filters cheaply before re-ranking with the model.

Live beta with active daily use across two organisations. Sub-second first-token latency at the edge.