I don’t know man… seems like an overkill. There are enough open source projects you could leverage and/or improve to bring it to production quality without reinventing the wheel.
I’d use self-hosted models on tiny + a few grad interns with occasional mentorship.
Put @comma_ai 's new Chestnut eGPU on my Comma 4 and ran the Lebowski model for an hour in my Rivian R1T — local roads + freeway. Is an $899 (was $799) GPU dock actually worth it?
youtu.be/UN2q7FXY4bg