I see these milestones:
Project setup (1 week; 1000€)
- provisioning of "AI" server (tilda won't do this because we need external Python packages for a flexible integration)
- integration from tilda to the AI server
- integration from AI server to OpenAI
- some initial prompt engineering for your use case
Fine-tuning (2 weeks; 1000€)
- data sources research
- preparing data and feeding it (we might use an OpenAI plugin or feed structured or unstructured data as "embeddings" to OpenAI; This will cost additional money, but is the best quality available currently). The data will be sent to OpenAI. You'll need to read their privacy terms to see if you are cool with how they are dealing with your data
- iterating on response quality and fine-tuning hallucinations
usage and user feedback tracking (optional; 1-2 weeks; 1000€)
- tracking of metrics and user input (this will be an extra integration; saving everything to a database and a separate dashboard to visualize it)
- additional performance and user monitoring (track response times and user reactions to it)
- user feedback button in tilda to see whether they liked the responses
I would publicly launch after the "usage and user feedback", but it can also be done before fully or as a soft launch.
After each milestone, there will be user testing and a discussion of the results. I would also appreciate having a product owner I could contact directly during the development phase.