This project is under continuous development.  If you want to get a head start, you can visit the repositories for this project at Git-Hub, and download the source codes that include the files from "Rubidiulm", "Propeller Debug Terminal", and the material from "Code Renaissance."  You will also need to have some kind of Microcontroller platform such as the Parallax Propeller P2 Eval or Edge Board, any FPGA-based system, or another environment that supports A-D and D-A conversion on multiple ports.

The training set used in the initial test case for this project, as of May 2023 is based on two compilations of some of the log entries created during the development of previous projects by this author, and that material is therefore available on this site.  Some have been made available in pdf form as the downloadable articles "Prometheus" and "The Money Bomb",

The current version of Megahal uses a training file entitled megahal.trn.  So if your "business" is "XYZ catering", you might want to write some kind of manifesto of customer support interactions and train off of that,  i.e., so that you should at least for now, be creating an ordinary mixture of text, source code, etc., in the "style" that you want the AI to learn.

In an earlier project, "Rubidium", I think - there was a demo of the application "Propeller Debug Terminal' which can be used to upload and interact with bespoke software that runs on any microcontroller that has a USB-compatible serial interface capable of UART emulation.  The Parallax P2 platform is specifically supported.

Other platforms should be relatively easy to adapt to.  For now, it will be sufficient to build an instance of MegaHal or Eliza, or any other classic AI, using any available 3rd party tool such as micro-python, cee-lisp, SPIN, Flex C, or Propeller assembly.  The built-in FORTH interpreter would also be an interesting choice. 

Based upon the project "Prometheus", a Pascal compiler derived from UCSD Pascal is also in the works but is not yet supported.  Likewise, the Frame LISP library contains some Pascal-style intrinsics for use with C/C++ programs and can be used to create a multi-threaded version of Mega-Hal, which in turn can run multiple models simultaneously, so that the training sets used can have multiple files associated with them, i..e. in a familiar makefile like format, using the scripts.txt file format.  This is actually how my old chatbot "Algernon" was written, and it is currently being revised for use with contemporary microcontrollers.

While we are not quite to the point yet of being able to give plain, natural language instructions to our computers, describing what we want them to do, the state of the art is getting very close.  Experiment to your heart's content.  The best results will most likely be obtained through an approach that combines techniques from traditional bespoke application development, with additional utilities and creative adjuncts that can be derived from so-called "deep learning".