To deliver LPython as an MVP, I think the bare minimum is:

• Create a few useful LPython packages, upload to PyPi and get working with pip and Conda naturally and easily (Compiling Python packages with LPython #992)
• @lpython decorator (Interface with Python (CPython) #703): get it working for arrays and simple types, and allow importing other LPython compiled packages from within it
• Add the @pythoncall decorator to create an "interface" for any Python function (in some Python module), similar how @ccall works, but for Python: Interface with Python (CPython) #703 (comment), Interface with Python (CPython) #703 (comment); this will be implemented by creating BindPython ABI (next to BindC), and then in the backends we implement the conversion; I would start with the C backend, where you add logic for BindPython (in FunctionCall ASR node) and generate C code that calls the Python C/API to import a given module, find the function, convert all arguments and call it
• Add @pythoncallable decorator which exposes an LPython function to CPython (similar to @ccallable)
• The demo using @lpython will be faster than other Python compilers (lists, dicts, some basic things like arrays and loops), this must be done via the C backend
• some LLVM backend demos: lists, dicts and arrays, anything where we are fast
• Add demo of using @pythoncall and @pythoncallable, via the C backend
• Website with documentation
• Blog post(s)