Why LLMs fail at gym training: deterministic rules vs. conversational AI

Hey Product Hunt community!

I’m Madhwesh, a software engineer building Ziddi, an AI-native human performance and wellness platform.

Over the past year of building, our biggest technical realization was simple: while generative AI is great at interpreting conversational context, pure LLMs should never prescribe heavy physical training. Hallucinating a workout variable isn’t just a bad user experience—it can lead to serious joint wear or real injury.

To solve this, we separated intent from execution:

• Using an agentic layer to interpret user goals and check real-time context.

• Backing it with hard, deterministic programmatic rules that act as safety gates—automatically overriding dangerous or contraindicated exercises based on user injury history.

For the builders and founders here experimenting with agentic workflows or healthtech:

How are you approaching deterministic guardrails vs. model flexibility in high-stakes user workflows? Where do you draw the line between generative reasoning and rigid system rules?

Would love to hear your thoughts and lessons learned!