I'm Samuel Castro
I'm a Computer Scienc student, passionate about technology and naturally curious. I love exploring how things work under the hood and I really enjoy developing Artificial Intelligence-related projects. My goal is to keep learning, solving complex problems, and creating useful and innovative tools.
Experience
Teaching Assistant
Universidad Nacional de Colombia - Sede Manizales
Served as Teaching Assistant for two consecutive semesters, supporting undergraduate students in foundational Computer Science courses.Programming I (2025-1): Designed and led hands-on workshop sessions, created practical exercises, and explained core programming concepts to help students build problem-solving skills from the ground up.Data Structures (2025-2): Provided one-on-one and group tutoring, resolving doubts on topics such as linked lists, trees, graphs, and algorithms.
Projects
A photo-sound journey through rural Colombian life. Web experience built for an anthropology project, featuring 10 audio postcards with over 35 minutes of field recordings capturing the stories, landscapes, and sounds of Anaime, Tolima, Colombia. Bilingual (Spanish/English) interface designed to preserve and share the oral and sonic heritage of the region.
A real-time system that detects news trends not by counting keywords, but by tracking the geometry of meaning. Every hour it ingests articles from Colombia's major outlets, embeds them with OpenAI text-embedding-3-small, clusters them with HDBSCAN on raw 1536-dim embeddings, computes a custom trend score combining burst, acceleration, and emergence signals, generates neutral multi-source summaries with GPT-4o-mini, and streams updates to the frontend in real time via SSE. Cluster identity is kept stable across cycles using centroid matching (cosine similarity ≥ 0.80).
Binary image classification system to detect AI-generated images vs. real photographs, addressing the growing threat of deepfakes and synthetic media. Uses Transfer Learning with a fine-tuned ResNet50 trained on the CIFAKE dataset (120,000 images), achieving 87.25% train accuracy and 84.90% validation accuracy. Deployed as a lightweight TFLite model (~25MB) and via Gradio demo on Hugging Face Spaces.
AI-powered web application for automated bone fracture detection in musculoskeletal radiographs, built to support radiologists in making faster and more accurate diagnoses. Uses a fine-tuned YOLOv8s model trained on the FracAtlas dataset (4,083 annotated X-ray images), achieving 95.5% precision and 94.5% mAP@0.5 — significantly outperforming the original paper's baseline. Users upload X-rays and receive instant fracture localization with bounding boxes and confidence scores.
Khipu is an AI-powered cybersecurity threat detection platform that democratizes access to advanced security analysis tools through Machine Learning and conversational AI. Combines 4 specialized ML models (spam classifier, phishing detector, suspicious access detector, network log analyzer) into a unified real-time dashboard with a natural language chat interface powered by OpenAI GPT and LangChain. Hackathon Winner.