Video search
Video Moment Finder
Open-source video search that combines transcript retrieval and visual embeddings so creators, researchers, and builders can find exact moments inside long videos.
React, Python, Node.js, semantic search
Projects
Selected personal projects, public experiments, and research notes across video, speech, retrieval, and music generation. Professional client and company work is not included here.
Music AI
Komposair
My first AI music project, built in 2020 before the current generative AI wave. Musicians enter a motif, then generate, save, and explore melodies using Magenta RNN models and custom Keras LSTM folk-song models.
Django, TensorFlow, Magenta.js, Tone.js, VexFlow
Music preservation
score2abc
A pipeline for converting handwritten Colombian music scores into ABC notation and metadata, with PDF rendering, staff segmentation, MusicXML/ABC export, QA, and evaluation artifacts.
Python, OMR adapters, ABC notation, preservation
View repoForecasting agents
Colombia Forecasting Desk
An experimental agent-assisted forecasting desk for Colombian political, economic, regulatory, and institutional events. It scans public sources, cleans and clusters items, tracks indicators, and produces daily forecast-candidate briefs.
Python, public-source monitoring, forecasting workflows
View repoImage AI
WhatsApp Image Bot
A FastAPI-powered WhatsApp bot for background and object removal, designed for online sellers and content creators. It connects Twilio, external AI image-processing APIs, cloud storage, and a commercial MVP direction around credits and payments.
FastAPI, Twilio, fal.ai, AWS S3, Docker
View repoMusic AI
Multitrack MIDI Music Generator
Interactive Hugging Face Space for generating multitrack MIDI by genre. The project turns symbolic music into model-friendly text, fine-tunes sequence models, and exposes the results through a playable demo.
Hugging Face Spaces, Transformers, symbolic music
Speech
Spanish Whisper Fine-Tuning
A set of Whisper fine-tuning experiments for Spanish automatic speech recognition, including Common Voice and FLEURS-style training runs.
Whisper, ASR, Spanish speech datasets
Dataset
LMD Clean 8-Bar Dataset
A cleaned version of the Lakh MIDI Dataset prepared for symbolic-music generation research, with 8-bar examples and train, validation, and test splits for model development.
345k rows, Parquet, Hugging Face Datasets
View datasetResearch notes
Diffusion from First Principles
A 52-week research curriculum that rebuilds diffusion models from probability foundations to diffusion, score-based models, SDEs, and flow matching, with notes, proofs, notebooks, and mini-projects.
Probability foundations, diffusion, SDEs, flow matching