We're building a new way to program a machine using just natural language - starting off with a fast way to build internal tools by describing them in English.
Hi, we are Sibesh and Shubham, founders of Maya Labs. We are on a mission to teach machines how to code - starting off with a fast way to build internal tools in minutes using just natural language. Here’s an early look :
A longer 5-min demo here :
Today, you can build one particular type of custom app well using Maya that let's you :
Maya incrementally generates programs & shows results in response to steps you type in English, so the whole experience doesn’t feel like debugging someone else’s code. It’s possible to modify, delete & debug the natural language parts and Maya responds instantly - which makes it possible to hold and move around the natural language program in your head. Checkout an early version of the docs here. We intend to keep it just one page long.
There’s still a lot to figure out and we’re trying to nail the experience for this one narrow use case first. When this works, we’ll try to extend it to just about any task you can think of - web navigation, connecting APIs, platform robots, RPA, home automation. Here’s a show reel of what else you can build manually with Maya today.
Who is this for?
Most internal tool builders are used by developers to build software for others on the product/ops/growth/sales/data teams, but we think this is a way for the user of the tool to also be the creator.
How does this work?
We use a combination of conditional logic, neural search/classification, fine-tuned language models and template generation to make this work predictably and reliably in response to words you type/modify. We generate programs in a custom flow-based programming language that lets us express a wide range of behaviours with a smaller program length and visual DAGs. We also created our own benchmark to measure how well machines are at generating code (see our paper on this here).
About us :
We met in college many years back while starting Hyperloop India, the first Indian team at the SpaceX Hyperloop Competition - and have worked on-and-off on various projects since, sharing a common interest for hard engineering problems. Program synthesis is the hardest of them all - if software is eating the world, and you could build software that could build other software, why work on anything else?
We’ve been tinkering on this for years. Many Python libraries, papers, toy environments and product iterations later, we think we finally have an angle of attack that has a chance of working.