enlight-ai Runtime Platform
AI architecture should feel like modelling cognition — not wiring APIs.
enlight sits between your app and your LLM. Requests flow through real DAG workflows — branching, caching, looping back through memory — not a linear chain of prompts.Self-hosted. Source available. Your data stays yours.
How it works
1 Expose
One endpoint or a hundred. Each one its own entry point, its own workflow, its own rules.
2 Configure
One model provider or ten. Mix Ollama, Anthropic, OpenAI, Gemini — assign them freely across your endpoints.
3 Design
Wire any workflow to any endpoint. Steps, tools, memory, branching — compose exactly what you need, nothing more.
Your topology. Your rules. No constraints.
Self-hosted, source available. Your models, your infra, your data. Nothing leaves your stack.
Real DAG workflows. Branching, memory loops, conditional exits. Not a pipeline.
No magic, no black box. Every step explicit, readable, testable. You own the code.
Ollama, Anthropic, OpenAI, Gemini. Swap per endpoint, mix freely.
Get started — or run a demo
⤷ RAG in one step
Your store, your retrieval logic.
Retrieve, inject, stream — one step, no framework lock-in.
⤷ Tool calling without magic
The model decides, you execute.
Fully explicit — no hidden protocols, every provider, every model.
⤷ Session memory that accumulates
Facts survive restarts.
The LLM always knows what the user told it — without you managing state.
⤷ Prompt enrichment before the main call
A silent pre-call extracts language, topic, and intent — the main call gets a richer context automatically.
⤷ Cognitive research pipeline
Four isolated steps: understand, search, reason, respond.
Grounded answers, source attribution, cross-turn memory.
⤷ Automated translation pipeline
Translate Markdown to multiple languages, validate section-by-section, open a GitHub PR with inline review comments.