For me, AI optimization early on is mostly about putting guardrails around response format and content so the output is consistent. I imagine that later it shifts more toward speed, cost, and reliability once the core workflow is stable.
What challenges are you solving right now with building an AI assistant?
Agree with this, that's the first level and the most accessible to everyone. There's so much more you can optimize when it comes to AI response that you can get analysis paralysis if you're not intentional.
it's a good idea, but it doesn't work in this use-case and would be against out privacy policy. We can't fine tune models for every tenant (they are thousands now and we just started). However we're trying building something custom using some adapters, prompting and tokenization.
Love to see AI solutions that help people live a happier life.
Assist bilingual conversations seems to be a gold nugget. My family is German-Hungarian. So I know the everyday struggles.
Claudia’s insights in building such a powerful product also speaks volumes.
Very cool, thanks for sharing!
Thank you for this awesome actionable framework.
Thank you for reading, Dennis :)
what's included in AI optimization?
We have a bit different approach building an AI Assistant for work (email/calendar/tasks) where focus is on optimization.
I think anyway it depends on the stage you're at at a certain moment.
For me, AI optimization early on is mostly about putting guardrails around response format and content so the output is consistent. I imagine that later it shifts more toward speed, cost, and reliability once the core workflow is stable.
What challenges are you solving right now with building an AI assistant?
Agree with this, that's the first level and the most accessible to everyone. There's so much more you can optimize when it comes to AI response that you can get analysis paralysis if you're not intentional.
the biggest challenge that we have is creating an email reply using tone of voice of the user.
tested about 200 variants of prompts, context... still not there 100%
Have you tried fine-tuning? That's more work but usually quite effective given enough samples
it's a good idea, but it doesn't work in this use-case and would be against out privacy policy. We can't fine tune models for every tenant (they are thousands now and we just started). However we're trying building something custom using some adapters, prompting and tokenization.
That's fascinating. What tools are you using?