ley-khaa
An assistant that turns a messy conversation into a finished job.
Personal product · 2026
Work rarely arrives as a clear instruction. It arrives as a conversation — half a request on Monday, a correction on Tuesday, a "actually, can it also…" on Wednesday. ley-khaa reads the whole thread, works out what was actually being asked, and turns it into one clear job that a computer can carry out.
Try it
Interactive demoTwo things, and the second is the hard one. It had to understand a scattered conversation well enough to be useful — and it had to know when not to act on its own. An assistant that quietly does the wrong thing at speed is worse than no assistant at all.
It reads the conversation and produces one plain-English summary of the job, which you confirm or correct before anything happens. Then it judges how confident it is and how risky the job is, and recommends how much freedom it should be given: suggest something, work alongside you, or go ahead. You always keep the final say.
Every job it completes comes with a record of exactly what it did and what it used, so any result can be checked or repeated later. In a business that matters more than speed.
Under the hoodFor the technical reader
Python and FastAPI behind a React front end, with Postgres, Alembic migrations and a Docker Compose stack that CI brings up on every push and drives through a seeded end-to-end demo. A cheap relevance filter runs before the expensive model call to keep cost down. The executor is synthesis-first: it writes a Python script for an open-ended task and runs it in a sandbox, with a workflow registry acting as a cache of proven fast paths. Every run emits a reproducible bundle — the deliverable, the generator code, the exact inputs and a seeded manifest — so a result can be audited and re-run. The backend is mypy-clean in CI, and the 1.0 release documents what is stable and what is not.
All data in the project is synthetic. It connects to no employer data, credentials or infrastructure.