Report 2. Initial Design & Elaboration #

UrTraining #

Authors: Ildar Rakiev, Makar Dyachenko, Salavat Faizullin, Egor Chernobrovkin, Alexandra Starikova, Ilona Dziurava, Anisya Kochetkova

Date: June 2025

Team Members #

Requirements #

User Story and Acceptance Criteria #

Backlog #

Design #

Core Application Design #

The core application design can be viewed here. These initial designs form the foundation of the core user experience. Below are the key screens that have been designed so far:

• User Questionnaire: A multi-step form where the user answers questions to receive a personalized training program.
  

• Prototype Review Window: Allows users to preview and modify their recommended training programs.
  

  

• Training Plan Catalog: Displays a list of available training programs when the user clicks on the “Catalog” tab in the top navigation bar.
  

• Training Card Overview: A detailed view of an individual training plan, including its description, difficulty level, and schedule.
  

User Flow Diagram #

User flow diagrams for key interactions can be found here: User flow diagram

Frontend #

During Week 2, we implemented several key pages and set up basic routing for the application:

• Welcome Page with Role Selection (User / Trainer)
  Implemented an entry screen that directs users based on their role.
  Welcome Page Commit

• Authentication Pages (Signup / Login)
  Added user authentication forms for registration and login, with working layout and form validation.
  Signup Commit
  Login Commit

• Basic Routing Setup
  Established initial routing logic to navigate between the main pages.
  Routing Commit

• User Questionnaire Page
  Created a multi-step form displayed after registration, allowing users to specify preferences for personalized training programs.
  Questionnaire Commit

Fronted part is deployed here.

Backend #

Progress during Week 2 was focused on defining the API contract and preparing the backend for upcoming integration with the frontend.

• Basic Endpoints with Dummy Data
  One to two base endpoints with dummy responses were already implemented last week.
  Dummy Data Endpoints Commit

• Core Business Logic and API Implementation
  Added initial backend logic and API endpoints to support MVP features.
  Core Logic Commit

• Database Integration
  Connected the project to the database and added a guide for working with it.
  Database Setup Commit

The following diagram illustrates the current architecture of our application:

Backend part is deployed here.

ML System Design #

Overview #

The ML system consists of two core components:

• Recommendation System: Delivers personalized course recommendations to users
• Image-to-Tracker: Converts exercise images into structured tracking programs

Recommendation System #

Current Approach #

Due to limited initial data, the system employs synthetic data generation for the first iteration model training.

Data Generation Pipeline #

User Data Structure #

• Profile Data: Registration form responses (weight, fitness goals, experience level, etc.)
• Interaction Metadata: User engagement metrics (completed courses, retention rates, activity patterns)

Synthetic Data Generation Process #

1. User Profile Generation

   • Multiple LLMs (DeepSeek, Gemma, LLaMA) generate diverse user profiles
   • Prompt engineering with seed phrases ensures profile diversity
   • Metadata generation aligned with corresponding user profiles
   • Target: 1,000 training examples for baseline model

2. Course Data Generation

   • Multiple LLMs generate the main content of exercises (for gym, swimming or other)
   • For each course type (gym, swimming, others) we generate separate training program
   • Finally, we align generated courses with coach-provided metadata forms (also generated by LLMs)

3. Dataset Assembly

   • For each user profile + metadata, DeepSeek R1 rearranges the courses (generates optimal course-user pairings for training data)

4. Model Training
   The training process utilizes a bi-encoder architecture to learn meaningful representations of users and courses. It aims to maximize similarity between compatible user-course pairs while minimizing similarity for incompatible ones, enabling effective personalized recommendations through vector similarity search.

   Architecture:

   • Bi-encoder models (e5, bge-m3, or similar embeddings models)
   • Separate encoding for user profiles and course descriptions
   • Vector generation for both user queries and course documents

   Training Objective:

   • Contrastive learning approach using triplet, margin, or contrastive loss functions
   • Optimize embeddings to place similar user-course pairs closer in vector space

Image-to-Tracker System #

Current Implementation #

The system utilizes open-source visual language models (VLM) to extract structured data from exercise images.

Process Flow:

1. Image Processing: Vision-language model analyzes input images
2. Text Extraction: Converts visual content to structured JSON format:

   [
     {
       "content": "exercise_description"
     }
   ]
   

3. Tracker Mapping: Each JSON object maps to a dedicated tracker page

Future Development (won’t be included in MVP) #

Data Collection Strategy: - Use images from real coaches to build the dataset - Annotate those images by DeepSeek or assessors - Training pair generation: (image) → (structured_tracker_json)

Users #

To better understand how trainers create and structure their training programs, we asked several of them to share real examples. This research helps us design a more intuitive and relevant upload feature that aligns with their actual workflows.

Below are four examples of training programs provided by different trainers:

It turned out that most trainers write their programs using tables, while one of them prefers to save and structure their programs in Telegram saved messages.

Weekly Progress Report #

During the second week, our team transitioned from planning to early implementation. The primary focus was on expanding the initial requirements, creating design prototypes, and setting up foundational structures for both the frontend and backend.

Challenges & Solutions #

One challenge we encountered was ensuring consistency between the frontend and backend specifications. We addressed this by closely aligning on the API contract early and scheduling regular sync meetings between the respective subteams.

Conclusions & Next Steps #

By the end of Week 2, we successfully transformed initial ideas into concrete requirements, established foundational designs, and set up project structures across the frontend and backend. The team now has a shared understanding of both the technical and functional aspects of the project, with wireframes, a prioritized backlog, and initial codebase scaffolding in place.

Looking ahead to Week 3, our focus will shift to MVP development, where we aim to implement core functionality and deliver at least one complete end-to-end user flow. This will involve connecting frontend components to backend endpoints, integrating with the database, and enabling meaningful interactions within the application.

Next Steps for Week 3 #

• Frontend: Connect UI components to real backend data and handle basic user interactions.
• Backend: Implement core business logic and enable database operations for MVP features.
• Database: Finalize schema and ensure it supports the initial set of use cases.
• Authentication: Begin integrating simple user auth flows.
• ML: Train a basic model version and expose an API endpoint.
• Testing & Demo Prep: Ensure the MVP flow is stable enough for an internal demo.

This upcoming sprint will be critical in turning the project from a prototype into a working product.

We confirm that the code in the main branch:

In working condition. Run via docker-compose.