Personal Project

Role: Solo developer (design, backend, AI, frontend, testing) · Type: Personal project Stack: Python · FastAPI · PostgreSQL · SQLAlchemy · OpenAI (GPT-4o, Whisper, TTS) · LangChain · Pinecone · Streamlit · pytest Code: <a href="http://github.com/AmerKone/Coachin" rel="noopener noreferrer nofollow" target="_blank">github.com/AmerKone/Coachin</a><h2>The problem</h2>Most fitness apps are either logbooks (you record everything, they explain nothing) or chatbots (they talk well but know nothing about you). And in a health context, a chatbot that improvises can do real harm: it can train someone through chest pain or suggest a crash diet.I wanted a coach that knows your data, can act on it hands-free, and is safe by design.<h2>What I built</h2>Coachin is an AI coach you can talk to during a workout. It:<ul><li>Builds personalized multi-week programs from a 102-exercise library, fitted to your equipment and injuries.</li><li>Tracks sets and suggests next session's weights with progressive overload.</li><li>Estimates meals from plain descriptions ("two eggs and toast").</li><li>Writes monthly progress reports with charts.</li><li>Answers questions about your history, e.g. "What was my bench press PR in January?", from your real logs.</li></ul><h2>Architecture</h2>A Streamlit frontend talks to a FastAPI backend (routes → services → SQLAlchemy models) over a JWT-authenticated REST API. PostgreSQL holds all user data across 14 tables.The coach is an LLM agent with 10 tools: log a set, log a meal, record weight, swap an exercise, look up history, and so on. A RAG pipeline (LangChain, OpenAI embeddings, Pinecone) grounds its answers in a curated knowledge base of training, nutrition and safety guidance.<h2>Key decisions and challenges</h2>1. Don't trust the model with hard constraints. When generating programs, I build the output schema per user, with only the exercises their equipment allows, so the model can't pick an invalid one. The plan is then validated in code, and the model gets one chance to fix any errors. Progression, difficulty caps and deload weeks are computed in code, where they're guaranteed.2. Safety that doesn't depend on the LLM. Every message passes a rule-based screen first. Emergencies (chest pain, fainting, suicidal thoughts), disordered eating and drug questions get fixed, pre-written responses, and the AI is never called. I tested the rules against 20 red-flag phrases and 15 look-alikes ("chest day", "dying to try deadlifts") with no misses and no false alarms. An LLM classifier handles subtler cases, and replies recommending unsafe calorie levels are blocked.3. Testing with realistic data exposed real bugs. I generated a full year of realistic history for a demo user, which surfaced issues that unit tests hadn't:<ul><li>A month ending on a lighter deload week looked like strength loss. The fix: exclude planned deloads from the trend.</li><li>Past months were judged against today's goals, so the report told someone in a fat-loss phase to eat more. The fix: use the goal and targets of that period.</li></ul>4. Knowing where LLMs are unreliable. The model mishandled date arithmetic ("last Monday"), so the prompt now includes a calendar of the past week, and the coach states the date it looked up. Prompt iteration with real outputs also caught the model copying a knee injury example into advice for a back injury.5. Honest charts. One axis per chart, a colorblind-safe palette validated by script, direct line labels, and a table view for every chart. I also rendered each chart as an image and reviewed it, which caught clipped axes and overlapping labels.<h2>Results</h2><ul><li>137 automated tests. Fake LLMs keep runs fast, free and deterministic, and every database test rolls back its data.</li><li>End-to-end checks with real models:<ul><li>A spoken meal description was transcribed word for word and logged by the coach.</li><li>Program generation respected injuries and equipment.</li><li>History questions matched the database exactly.</li></ul></li></ul><h2>What I'd do next</h2><ul><li>Cut the coach's response time (currently 10–34 s per voice turn) by running the safety check and knowledge search in parallel and streaming the spoken reply.</li><li>Harden it for real users: login rate limiting, and a separate test database.<div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmupmozvx0xd24kd0c1w56gc1.png"><div class="image-overlay"></div></div></div></div></li></ul>

An AI fitness coach you can talk to. Coachin builds personalized training programs, tracks your workouts and meals, suggests your next weights with progressive overload, writes monthly progress reports with charts, and screens every message for medical red flags.

Built with FastAPI, Streamlit, PostgreSQL, OpenAI (GPT-4o, Whisper, TTS) and Pinecone.

Ameur Kaouane

Ingénieur GenAI / IA — RAG & LLM

À propos

Formation

Université Évry Paris-Saclay

École Supérieure en Informatique de Sidi Bel Abbès – ESI SBA

Licences et certifications

AWS Certified ML Engineer – Associate

Portfolio

Coachin

Personal Project

Talents

Recruteurs

À propos