ML engineer wanted: self-host + fine-tune a Japanese LLM for structured tasks

We have an AI assistant feature in a web product, currently running on a commercial LLM API. It classifies user requests, triggers UI actions, fills forms from natural language, and answers questions from page content — in English and Japanese. We want to replace the external API with our own self-hosted, fine-tuned open model. Main reasons: cost control and data privacy (our Japanese users prefer data not going to third-party AI providers). What we need: • Pick the right open model — strong Japanese is a must. Our workload is mostly structured decisions and JSON output, not long-form writing, so we suspect a smaller model is enough — confirm or challenge that. • Set up hosting and give us honest monthly GPU cost numbers first — budget not yet fixed, we want a cheaper option and an ideal option. • Fine-tune it on our tasks — we provide the training data. • Prove it works — we have a test set recording how our current setup answers 100+ real queries. Your model must match it in both languages before we switch. • Stay on after launch for monitoring, retraining, and server upkeep — quote a monthly retainer. Tech context: Python backend with a retrieval pipeline and vector database, on cloud infrastructure in the Tokyo region. Full details shared after shortlisting under a mutual NDA. You'll coordinate with our existing developer, and should be comfortable with AI coding tools. I'm non-technical, so plain-language communication is important. To apply, skip the template. Just answer: (1) what model size do tasks like ours really need, and realistic monthly hosting cost? (2) how do you prove the new model matches our current quality before switching? Real answers get interviews — specifics shared under NDA.