1
1. You’re building a airplane ticket booking platform, and want to offer users forecasts of when ticket prices will go up or down in the next 30 days.
2. You have a retail shop, and want to build an internal dashboard that forecasts how many items of each type you need to buy next week, given how many you sold in the last month
3. You’re building an real estate investment platform, and want it to notify users how real estate prices will change over the next 5 years (given sq. m., distance, location, etc.)
Problem:
You want to add forecasting to the app or platform you’re building
However, you don’t know enough to build and deploy a production-grade ML model
What if you could have a platform that could autonomously train and deploy ML models for you?
Solution:
We propose a vibe coding platform (like https://lovable.dev/ )
To empower everyday users to train and deploy machine learning models to production without needing to know anything about the field
Should we build it?
Building in public is important for us, and we wanted to hear your take
We know that payments are an important part of agentic tool usage
However, right now, UTCP doesn't offer any possibility for the agent to use paid tools without human assistance.
So agents can’t self-provision/pay for API keys: humans do it and pass the token in.
So the question to you all is: Should UTCP define an interface/set a standard for third-party payments/getting auth so agents can safely obtain keys or pay per call via plugins.
And if so, how would that look like? How would you like the agents to be able to pay and how would you like your tools to be able to monetize themselves?
Options: 1. Stay out: leave payments/API keys to each tool and client to figure out, no standardization through the protocol 2. Define a way for tools to advertise how they can get paid through the manual, so agents can pay if the user approves: give us ideas on how you think this could be done, and how you would like it most
Goal: make paid tool usage practical for agents w/out human needing to go through a long process of getting the API key for everything. What are your thoughts?
I've been diving into tool-calling with some local models and honestly, it's been a bit of a grind. It feels like getting consistent, reliable tool use out of local models is a real challenge.
What is your experience?
Personally, I'm running into issues like models either not calling the right tool, or calling it correctly but then returning plain text instead of a properly formatted tool call.
It's frustrating when you know your prompting is solid because it works flawlessly with something like an OpenAI model.
I'm curious to hear about your experiences. What are your biggest headaches with tool-calling?
What models have you found to be surprisingly good (or bad) at it?
Are there any specific prompting techniques or libraries that have made a difference for you?
Is it just a matter of using specialized function-calling models?
How much does the client or inference engine impact success?
Just looking to hear experiences to see if it's worth the investment to build something that makes this easier for people!