EL CRUCE
PLANNER
Built during an LLM Engineering course, then actually used: training for and completing the El Cruce ultramarathon in Patagonia, 2025.
A HYBRID PIPELINE
Not a single model doing everything. Each piece does the job it's best at.
Deterministic core
Geospatial data, pacing, and fatigue calculation. The part that shouldn't be left to a model.
Local coaching layer
A local LLM for the coaching voice — private, offline-capable, running on the machine.
Live data & UI
Strava API pulls real training data in. Streamlit + Plotly tie the system together for visualization.
IN MOTION
NO LANGCHAIN
Deliberately skipped LangChain. The architecture didn't need the overhead, and lean held up better than following the framework of the moment.
The only way to actually learn a new technology is to build something you need with it, then use what you built.
BUILT TO BE USED
Course project → race-day tool. Patagonia, 2025.