A mobile-first gym workout tracker with built-in AI coaching, self-hosted on your own AWS account. No SaaS, no subscriptions, no one else holding your training history.
Set, reps, weight, RPE — with a rest timer and per-exercise history right next to the input.
Run full-body, push/pull/legs, or whatever split fits the week. Activate one plan, keep the rest as drafts.
Per-session debriefs and weekly volume reviews. Pick the model: Bedrock, Gemini, OpenAI, Anthropic, or off.
Weekly muscle-group coverage, volume and intensity trends, a 2-week activity heatmap.
Drill into any past session, edit a stray set, or open an exercise to see every rep you've ever logged for it.
Italian and English, switchable from the profile. AI coach replies follow the same language setting.
Schedule the week, then let the app flag when you've stacked too many sets on one muscle group. The 14-set chest day above lights up because it's over the configured limit — tunable per muscle from your profile.
Every past session shows up with the muscles you actually worked rendered on a body map. Sets, total kg-rep volume, average RPE — enough to know how the week went without opening a single record.
The weekly review reads your real volume against the plan and writes back specific moves: "add 12 back sets, split between Monday and Wednesday, here's the breakdown." Same prompt across Bedrock, Gemini, OpenAI, or Anthropic.
Your athlete description and free-form notes are pasted verbatim into every coach prompt. Tell it once that you're young, yout gender, your goal and how to exclude exercises — every suggestion afterwards respects that.
The project is a personal one, but the decisions are documented. Here are the ones that aren't obvious from the README.
Next.js server actions talk to DynamoDB through an IAM role bound to the Amplify SSR runtime. One layer less than the typical AWS web stack, same auth boundary — the browser still only sees Next.js endpoints behind a Cognito-validated cookie.
A single generateText dispatcher sits behind four backends.
Swap with one env var; prompts stay portable. The coach was rebuilt
against Bedrock, Gemini, OpenAI and Anthropic without touching the
prompt code.
A tiny dictionary module typed against a single Dictionary
interface — both locales must compile, so a missing translation breaks
the build, not production. No URL prefixes, no SSR/CSR mismatches,
no runtime negotiation in the hot path.
The Profile page holds an athlete description and a free-form notes field. Both are pasted verbatim into every AI coach prompt — so the suggestions account for your age, goals, and quirks ("no squats, bad knee") without you re-explaining them every week. No forking, no env-var sprawl: the same row drives the rest of the app.
Clone your fork locally. The infra and the web app live in the same tree.
sam deploy --guidedOne SAM template provisions a Cognito user pool, 5 DynamoDB tables, the runtime IAM policy, and the compute role for Amplify SSR.
Point Amplify at your fork, paste the six env vars from the SAM outputs, bind the compute role, deploy.
/setupFirst login on an empty stack lands on the setup wizard: two clicks seed ~27 default exercises (in your language) and a 2-day starter plan you can edit from there.
Read the full guide → Every step is documented in the repo README — IAM bindings, env vars, the lot.