I like your point about software for ML, vs software for decision making. I reckon there is a lot of existing software for decision making, but it is focused around particular domains/industries. A few months ago there was some news about a new startup offering experimental design as a service -- that's another related, seemingly under-explored idea.
Branch & bound solvers (aka combinatorial optimisation) as a service doesn't sound crazy to me, but perhaps one would need to think very carefully about the market.
What kind of customers would this service have?
If you are aiming to win customers by offering a cheaper service than alternatives, then I suppose the individual customers would need to be only the ones where it made more sense to rent this infrastructure instead of invest in their own. I.e., they would need some usage, but not usage heavy enough to justify investment in their own infrastructure, which would be cheaper for them in the long run.
Might have to think about data-security issues too, if commercially sensitive data is being uploaded to the solver back-ends.
Context:
I have worked somewhere that internally runs a service vaguely similar to what you describe. E.g. licensed commercial solver, sitting on a server with a decent amount of memory and compute, used as a back-end by various services to solve sufficiently valuable business problems for clients.
If one built a service like this, another idea is to keep it to yourself, and partner with some operations research consultants, then go directly after the business problems.
Branch & bound solvers (aka combinatorial optimisation) as a service doesn't sound crazy to me, but perhaps one would need to think very carefully about the market.
What kind of customers would this service have?
If you are aiming to win customers by offering a cheaper service than alternatives, then I suppose the individual customers would need to be only the ones where it made more sense to rent this infrastructure instead of invest in their own. I.e., they would need some usage, but not usage heavy enough to justify investment in their own infrastructure, which would be cheaper for them in the long run.
Might have to think about data-security issues too, if commercially sensitive data is being uploaded to the solver back-ends.
Context:
I have worked somewhere that internally runs a service vaguely similar to what you describe. E.g. licensed commercial solver, sitting on a server with a decent amount of memory and compute, used as a back-end by various services to solve sufficiently valuable business problems for clients.
If one built a service like this, another idea is to keep it to yourself, and partner with some operations research consultants, then go directly after the business problems.