Megh Vyas
Backend · Infra · SRE
← All work

AI workflow compiler

24LLM

Describe an automation in a sentence. Get back an editable, tool-wired node graph that actually runs it.

24llm.com

The compiling problem

Turning an automation idea into a working pipeline usually means hand-wiring nodes, picking a model per step, and hoping it behaves the same way twice. Most 'no-code' builders start you on an empty canvas and leave the compiling, the model choices, and the guardrails entirely up to you.

The compiler

01

Built a natural-language-to-workflow compiler that parses a plain-English description straight into an editable, tool-wired node graph instead of a fixed template.

02

Wired per-node model selection, so each step in a workflow runs on the model suited to it rather than one model for the whole pipeline.

03

Added human-approval steps as first-class nodes, so a workflow can pause for a person before it continues.

04

Instrumented full execution tracing across the graph, so every run is inspectable step by step, not just pass or fail.

05

Built the backend and compiler end-to-end: the language-to-graph translation, the execution engine, and the API the editor runs on.

Sentence to system

A sentence goes in; a working, editable automation comes out, already wired to the right tools and models, with a human in the loop wherever the workflow needs one.

0

sentence to workflow

per-node

model selection

full

execution tracing

Wired in

Natural-language-to-workflow compiler
Editable, tool-wired node graph output
Per-node model selection
Human-approval steps as first-class nodes
Full execution tracing
Backend and compiler built end-to-end