Tanda Runner: A personalized running dashboard

Tanda Runner is a web app that shows me the things I care most about when preparing my next marathon. I've also added a running coach/agent designed to give me actionable feedback about my training. Some of that feedback is probably hallucinatory - I'll get to it in a bit.

It's built with Django - a framework I've gravitated to more and more when building these types of apps. It has all the batteries I need and a time-tested ecosystem. The front-end is a Frankenstein that uses Django channels, htmx, and some vanilla JavaScript. It’s not the most responsive PWA out there - but it's still a pretty usable experience. The chat interface is adapted from LLaMA.cpp.

The most interesting part is the small LLM coach/agent that gives feedback and tips regarding my performance. It's an untested and non-evaluated multi step LLM pipeline that extracts running-related insights from around the web, and uses them to give me personalized feedback using my Strava data.

To do this, I extract transcripts from marathon related YouTube videos (which I probably shouldn't), and use those to generate a list of insights related to running and training. Once I have those insights, I use an LLM to translate those them into Pandas code. Once that's done I can feed an LLM the insight, the code, and the outcome of running such code on my own data. Sounds confusing? I know it does - but it works surprisingly well!

This pipeline could probably be significantly improved by designing some evaluations, some metrics, and continuously improving it. But for now, I'll use it to prepare Nice, and make any adjustments I see fit.

The visualizations and graphs are built using Altair. They are a mix of my favorite graphs from Christoph’s CR Plots, Strava, and my own brain. The name (and most of the visualizations) come from Giovanni Tanda's work - which I encourage you to read through!