Tim Le

myGymBro

An AI fitness coach that combines your workout logs, a 132-source RAG knowledge base, and next-weight prediction models.

Role
Solo developer
Timeline
Apr 2026 – Present
Team
Solo
  • Python
  • Streamlit
  • RAG
  • Sentence Transformers
  • XGBoost
  • LiteLLM

TL;DR

The problem

Most workout trackers are good at storing sets and reps but give no actionable coaching. Generic chatbots give advice, but they don't know what you lifted last week.

myGymBro joins the two: user state (logs, profile, goals, equipment), fitness knowledge (RAG), and predictive modeling (what weight should I try next?).

How it works

  1. 1

    Log

    Workout Logger page captures exercise, sets, reps, weight, RPE and RIR into a per-session store.

  2. 2

    Profile

    Goal, experience level, equipment access and weekly frequency become structured training context.

  3. 3

    Retrieve

    The question is embedded with all-MiniLM-L6-v2; top chunks are re-ranked by metadata tags, trusted domains and low-value-text filters.

  4. 4

    Predict

    For progression questions, a hybrid of the baseline and XGBoost models estimates the next working weight.

  5. 5

    Generate

    Retrieved context, recent history and predictions are sent as system context to the LLM (via LiteLLM); the reply is streamed back to the chat.

Tech stack & why

App

  • StreamlitFast multi-page UI for a data-heavy app without a separate frontend.
  • LiteLLMOne interface across model providers; easy to swap models.

ML

  • Sentence TransformersLightweight embeddings that run cheaply for retrieval.
  • XGBoostStrong tabular baseline for next-weight regression.
  • Pandas / scikit-learnCleaning, feature building and evaluation.

Hard parts & key decisions

Making the coach feel integrated

Passing retrieved context as system-level app context (instead of pasting it into the user message) stopped replies like "based on the context you provided" and made the assistant feel like part of the app.

When the simple model wins

The rule-based baseline had the best MAE and within-2.5/5/10 lb accuracy because real lifters progress in discrete plate jumps; XGBoost had the lowest RMSE. The shipped app picks between them based on how much history a user has logged.

Noisy retrieval

YouTube transcripts are full of filler. Re-ranking on trusted domains and down-weighting phrases like "thanks for watching" noticeably improved the chunks the coach sees.

Results

132
knowledge-base sources
721
training workouts
420
held-out test rows
3
models compared

What I'd do next