Francisco De Haro

Repai Trainer

Live

Repai Trainer is a personalized AI-powered training mobile application, available on the App Store. Fully developed by me from scratch, it combines a native iOS frontend with Expo/React Native, a robust Supabase backend, subscription system with RevenueCat, and workout generation using GPT-5. The app allows users to create personalized routines, track their progress, and receive intelligent exercise recommendations based on their goals and performance.

Personal project currently on development pause.

UMChat - Advanced Chat System

Live

UMChat is an intelligent chat web application built from scratch with a complete system for managing conversations, documents and authentication. The project demonstrates expertise in modern full-stack development, complex data architecture and performance optimization. Includes AI integration, automatic OCR, intelligent file deduplication, real-time streaming, and an architecture of 8 tables with complex relationships. Built with Next.js 14.2, React 18.3, TypeScript, PostgreSQL with Supabase, and multiple external APIs.

Personal project representing the evolution of Testum towards a complete conversational platform.

Testum

Live

Testum is a web platform developed with Next.js that allows creating and taking multiple-choice exams quickly and effectively. Users can upload their own exams and solve them with a highly polished UX experience. The project reached +1600 questions to solve and +21400 accumulated answers, spread solely by word of mouth among students of the Business Administration degree at my university. It includes an integrated AI chat with OpenAI to help users while answering, authentication with Supabase, and PostgreSQL database to store all information.

Personal project developed for Business Administration students at my university.

Demo Mistral Chat

Live

A modern chat web application that demonstrates Mistral AI capabilities. Built with Next.js 14 and deployed on Vercel Edge, it implements real-time response streaming for a fluid and responsive user experience. The project explores best practices for integrating LLMs into web applications, including context handling, prompt optimization, and efficient conversation state management.

Llama Fine-tuning for Spanish Lyrics

Archived

Fine-tuning project of Llama models (3.3 1B/3B and 3.1 8B) to produce coherent Spanish song lyrics with consistent structure, rhyme, and theme. Optimized to balance quality and cost through advanced techniques like QLoRA 4-bit and Unsloth. The project includes a complete custom scraping pipeline developed with Playwright to create a robust dataset of +10,000 prompt-completion examples, intelligent tokenization system with masking, and efficient training architecture with A100 achieving high-quality results with ~24.3M trainable parameters.

Personal experimentation project with LLM fine-tuning. Training performed with A100 GPU.

LoRA Black Stories - Image Generation

Archived

LoRA adapter training project to replicate the characteristic visual aesthetic of "Black Stories" and generate new stories/scenes with visual consistency. The project includes meticulous dataset curation, autocaptioning system to strengthen style, trigger word definition for precise style control through prompts, and multi-resolution training (512-1024) to improve generalization. It includes a rigorous aesthetic QA process reviewing composition, silhouettes, contrast, contours, and coherence between samples to ensure stable and reproducible aesthetics. Trained on Black Forest Labs FLUX models (flux-schnell and flux-dev).

Personal project for own use in image generation and style control experimentation.

1¢, 1min, 100%: AI in University Exams (In Development)

Dev

Opensource research paper on the application of artificial intelligence models to Business Administration and Management exams at the University of Murcia, Spain. Results and methodology are publicly available. Results demonstrate that advanced reasoning models like GPT-5-mini (high reasoning) and grok-4-fast-reasoning achieve 100% accuracy in less than 1 minute per exam, with an approximate cost of 1 cent per evaluation. This study includes the development of an OCR system implemented with Mathpix API and LATEX/Markdown conversion program to automate exam processing. SQLite database with Prisma for efficient data management.

Academic research in development. Results are preliminary and subject to review.