Projects
Case studies
Other projects
P05: Futurisys, Employee Attrition Prediction
Problem. Develop a machine learning tool for Technova Partners' HR to identify employees at high risk of attrition using various professional parameters.
Approach. Applied MLOps principles to deploy a Random Forest Classifier (17 features, 5 engineered) using FastAPI, PostgreSQL for tracking, and a cloud CI/CD pipeline.
Result. The model reaches 83% recall and 0.35 precision on the test set.
scikit-learn FastAPI PostgreSQL Docker
Repository
P06: Credit Scoring, Model Selection and Optimization
Problem. Develop a credit scoring tool for "Prêt à dépenser" to calculate loan repayment probability and classify applications as either approved or denied.
Approach. Developed classification models using cross-validation and Optuna for hyperparameter tuning, optimized the business threshold, and implemented model registry and serving using MLflow. The champion model is the one deployed in P08.
Result. The optimized LightGBM model minimized the business cost to 29,778.2 across 5-fold cross-validation, achieving an average AUC of 0.79 with a 0.48 threshold.
scikit-learn LightGBM MLflow Optuna SHAP
Repository
P07: Puls-Events, Cultural Event Recommendation Assistant
Problem. Let a user ask a question in plain language and receive cultural event recommendations grounded in real listings rather than generated from the model's own knowledge.
Approach. Built a RAG system over Open Agenda event data: a FAISS vector index for retrieval, an LLM (Mistral or Gemini) for contextualised answers, and a FastAPI REST service exposing it to business teams.
Result. Automated Ragas evaluation: context precision 1.000 and answer relevancy 0.807, against faithfulness 0.624 and context recall 0.623. Retrieval returns relevant passages but misses part of the ground truth, and generation does not stay fully anchored to what it retrieves.
LangChain FAISS Ragas FastAPI
Repository
P08: Credit Scoring, Deployment and Monitoring
Problem. Deploy a credit default scoring model (LightGBM, 20 features) into production as a reliable and accessible API on Hugging Face Spaces.
Approach. Productionizing the LightGBM model built in P06. Implemented a CI/CD pipeline (Docker, Hugging Face) with a 0.48 business threshold, integrating Evidently for data drift monitoring.
Result. 94% test coverage, ~253 ms end-to-end inference, ~896 ms model loading.
gradio joblib evidently pytest
Repository
P10: BrainScan AI, Semi-Supervised Brain Tumor Detection
Problem. Develop a semi-supervised pipeline to classify 1,506 brain MRI scans where only 100 images are initially labeled as normal or cancerous.
Approach. Extracted ResNet18 features, applied PCA and K-Means for weak labeling, then performed two-phase semi-supervised training to outperform a standard supervised baseline.
Result. 80% accuracy on the test set against 55% for the supervised baseline.
torch torchvision scikit-learn opencv-python
Repository
P11: Eagle-1, Autonomous Lunar Lander Agent
Problem. Develop a reinforcement learning agent to safely and efficiently land a spacecraft in the LunarLander-v3 environment, and deploy it for user interaction.
Approach. Trained a PPO reinforcement learning agent, exposing the inference model via a FastAPI backend connected to an interactive Streamlit dashboard.
Result. Average score of 278.56 (+/- 17.76) over 100 deterministic evaluation episodes.
stable-baselines3 gymnasium fastapi streamlit
Repository
P12: Multimodal ETL Pipeline for Fake News Detection
Problem. Continuously supply a fake news detector with an up-to-date, structured corpus that accurately links news text to their corresponding images.
Approach. Built an automated text and image acquisition pipeline orchestrated by Airflow, which ingests RSS feeds, cleans data, and loads it into PostgreSQL.
Result. On a second consecutive run, no errors and no duplicates: 21 of 22 entries were already present and were updated.
Apache Airflow PostgreSQL Docker Compose pandas
Repository