Virtual Trainer

Problem

My co-founder and I wanted a trainer, not a tracker. Most fitness apps we tried logged workouts well enough but planned them badly, made you fight the screen mid-set, or hid your progress. We wanted an app you could trust to plan your training, that stayed quick between sets, and that still made sense after months of use.

Approach

I led the product and built the app and the back end on my own. The React Native client is built around a model of how training actually works: routines, workouts, exercises, timers, warmups, supersets and history. Each user arranges their own dashboard, and its results are cached per user, so it stays quick as people add more. Guided workouts progress automatically, using the same rules of thumb a coach would. Later I added an AI trainer that answers from the user's own training data instead of giving generic fitness advice. Most of the work there is choosing what goes into the model's context for each question: what's relevant, what's out of date, and what's too sensitive to send at all. If the trainer suggests a change, like swapping the bench press for dumbbells, the change goes through the same checks as one made in the app before it touches real data.

The Virtual Trainer dashboard with charts for workouts completed, weekly streak, workout frequency and volume. — The dashboard

A guided session on Bench Press: 4 sets of 10 reps at 150 lbs with 1:15 rest, and a note from the trainer on where you left off. — A guided set

Bench Press in the exercise library, with its demo, the member's personal records and step-by-step instructions. — The exercise library

Before building a plan, the trainer asks whether anything has changed with your health, with the answers Nothing new and Something's changed. — Planning a routine

After a member asks to swap the bench press for dumbbells, the trainer confirms the swap and the Upper Body day shows Bench Press with dumbbells. — Asking for a change

A member asks the trainer how to breathe during a heavy squat, and the trainer answers with a link to the squat walkthrough. — Asking the trainer

Architecture

The apps talk to a C# .NET API, 86 endpoints across 14 controllers, split along the same lines as the training model. On Azure, SQL Server holds workouts and accounts, Cosmos DB holds the data that is naturally document-shaped, and Service Bus carries background work like notifications so it never holds up a request. In the app I kept screen state, shared app state and server data apart, which kept long workout sessions responsive.

Tradeoffs

Automatic progression needed limits so it could adapt without suggesting unsafe jumps. Running both SQL Server and Cosmos DB costs more to operate, so I kept a clear rule for what lives where. Moving work to the background made the app faster, but anything the user needs to see right away stays out of the queue.

Outcome

Virtual Trainer shipped on the App Store and Google Play with generated routines and workouts, guided training, progress reports, custom exercises, automatic progression and the AI trainer. I designed and ran the whole stack in production, from the mobile apps down to the cloud. The app and the API carry more than 6,000 automated tests between them, which is what keeps it safe to change.