Hi everyone,
I’ve been experimenting with a personal training agent called Liv, built around Intervals.icu as the core planning system.
Liv is not a generic AI chatbot.
She’s an agent with an explicit persona, strict context handling, and defined decision boundaries.
What the agent does
- Interprets athlete bio, current physiological status, and season goals
- Performs macro planning (season → phase → week), not isolated weekly optimization
- Actively constrains training decisions (e.g. recovery protection, load progression)
- Can reject or counter athlete intent if it violates the plan logic
The training concepts and planning heuristics are largely based on established endurance frameworks, inspired by coaches like Joe Friel.
Technology / data stack
- WHOOP → readiness & recovery signals
- Intervals.icu → training history, load metrics, calendar context
- n8n → orchestration, state handling, data flow
- OpenRouter → controlled LLM usage and model comparison
- Telegram → conversational interface
Architectural note
The agent handles status interpretation, macro planning, and weekly planning.
The workout builder itself is fully deterministic — no AI-generated intervals, no stochastic session structure.
AI is explicitly excluded from low-level workout construction.
I shared a short Instagram video showing how a training week is derived and adjusted in practice.
Happy to discuss.
