Daniel Darquea

Cocoa Disease Localization
YOLO + MLflow

End-to-end pipeline for detecting and localizing cocoa plant diseases using YOLO, MLflow experiment tracking, and OpenVINO optimization, deployed as a Gradio app on Hugging Face Spaces.

Actions

• Reviewed and corrected annotations using Label Studio.
• Logged exploratory and full training runs with MLflow on DagsHub.
• Benchmarked inference time and accuracy across model formats.
• Built and deployed a Gradio interface on Hugging Face Spaces for interactive testing.

Techniques

• YOLO11n for object detection
• MLflow + DagsHub for experiment tracking and model registry
• ONNX and OpenVINO export & validation
• Performance benchmarking (accuracy vs inference time)
• Gradio + Hugging Face Spaces deployment

• YOLO11n
• MLflow
• OpenVINO
• Computer Vision
• Deployment

Vehicle Tracking and Speed Estimation using Computer Vision.

Detection, tracking, counting, and speed estimation of cars from video sequences using deep learning and classical computer vision techniques.

Actions

• Applied YOLO11n for vehicle detection and tracking across frames.
• Extracted road region with Canny edge detection and Hough line transform.
• Applied perspective transformation to generate a bird’s-eye view of the road.
• Defined counting lines to measure vehicle flow in both directions.
• Estimated vehicle speeds from tracked motion in the bird’s-eye view.
• Created annotated output video with bounding boxes and counts.

Techniques

• YOLO11n for object detection
• OpenCV: Canny, HoughLinesP, perspective transform, VideoWriter
• Vehicle tracking with ID persistence
• Geometric transformations for speed estimation
• Custom visualization pipeline for results

• YOLO11n
• OpenCV
• Computer Vision

Detection of Obstructive Sleep Apnea using Deep Learning on Audio Signals. Part 1: Tracheal Microphone Data

Detection of obstructive apnea events from tracheal audio recordings during polysomnography.

Actions

• Extraction and preprocessing of tracheal microphone signals from PSG .edf files.
• Conversion of 10s audio segments into normalized Mel spectrograms.
• Training EfficientNetv2B1 on balanced Apnea/No Apnea samples.

Techniques

• Custom preprocessing pipeline with TFRecords.
• Data balancing per patient and event type.
• Mixed precision training and Cosine annealing Scheduler
• Definition of adequate Losses, Metrics and Results Analysis

• Tensorflow
• EfficientNetv2B1

Building segmentation from satellite imagery

Building segmentation using satellite imagery

Actions

• U-Net model trained on Google's TPU for building segmentation.

Techniques

• Data Selection
• Fine Tuning
• Definition of adequate Losses, Metrics and Architecure

• Tensorflow
• seresnext50 encoder U-Net

Rocket League Game Score Prediction

Predict team scores in Rocket League within 10 seconds

Actions

• Built a deep learning model for score prediction.

Techniques

• Used the adequate model architecture and applied cross-validation techniques to optimize predictive accuracy of the Deep Neural Network proposed. Engineered features impact the most over final estimations.

• Tensorflow
• Binary Classification

Sales Forecast for Large Scale Bookstore

Predict future sales using time series forecasting

Actions

• Developed a composed forecasting model using store, product rates, and seasonality data models.

Techniques

• Implemented Ridge Regression, ARIMA and other data models performed by Exploratory Data Analysis (EDA) to identify weekend, holiday, and seasonal effects.

• ARIMA
• Ridge
• Composed Model

California Housing Price Prediction

Predict median house prices based on property features

Actions

• Engineered features and created a regression model for house price prediction.

Techniques

• Applied a combination of Regression Models (XGBoost, LGBM and CatBoost) with cross-validation to improve model accuracy.

• Catboost
• XGBoost
• LGBM

DATATON 2023 BANCOLOMBIA: Linear Optimization Problem

Schedule Workers Optimization Problem

Actions

• Set equations describing the problem to optimize.

Techniques

• Applied Linear Optimization methods to find the most efficient worker schedule.
• Developed scenario-based decision-making methodologies.

• Python
• PuLP

Santa 2022 : The Christmas Card Conundrum Challenge

Linear programming model minimizing logistics cost across multiple warehouses.

Optimize a traveling salesman problem for image printing.

Actions

• Solved a complex optimization problem to find the best route for printing an image pixel by pixel.

Techniques

• Set configuration constraint to simplify optimization.
• Used Traveling Salesman algorithms (LKH).

• NumPy
• LKH