Innovative AI solutions from research to production. Exploring the intersection of machine learning, finance, and practical applications.
Reinforcement learning system training AI to play Briscola (Italian trick-taking card game) via self-play against an LLM opponent. Custom PPO implementation with 328-dim observation vector, action masking, and margin-based reward shaping. Full-stack web interface for human vs AI gameplay.
AI-powered content generation platform for TikTok scripts, carousel posts, and animated social media content. NestJS backend with Claude Opus 4.6 for intelligent content creation, Playwright for server-side rendering, and Angular 21 frontend. Docker Compose deployment with GitHub Actions CI/CD.
Production AI-powered diet generator with weekly meal plans (35 meals/week), AI recipe generation, and smart shopping lists. Live at diet.silviobaratto.com with Google authentication and subscription system. Complete full-stack solution with Docker deployment.
ML system forecasting F1 World Championship outcomes using official telemetry data. Combines Bayesian variance decomposition, 10,000+ Monte Carlo simulations, and ensemble ML (Random Forest, Gradient Boosting, SVM). Features 50+ engineered features, track classification, and hybrid predictions with confidence intervals.
AI chatbot for Italian regions with scraped knowledge base and real-time SSE streaming. Per-region knowledge (starting with Friuli Venezia Giulia), RAG pipeline with Qdrant vector search, points of interest bookmarking, and conversational AI powered by GPT-5 Mini. Live at italia.silviobaratto.com.
Privacy-focused RAG system running entirely locally. Supports PDF, Word, PowerPoint, Excel, Markdown, and HTML files. Features semantic chunking, interactive CLI with model selection, dual modes (direct chat or document-augmented), and streaming pipeline for memory optimization.
Advanced Conditional VAE generating 64×64 Pokemon images from text using dual conditioning: CLIP ViT-L/14 (768-dim) + categorical attributes (type, color, shape). Features BAML metadata pipeline with OpenAI vision, LPIPS perceptual loss, and 17.8M parameters. Production-ready with FastAPI + Angular frontend.
Quantitative portfolio construction platform built on skfolio and scikit-learn. Features 13 optimization models, HMM regime blending, walk-forward cross-validation, vine copula scenario generation, and 8 investability screens. Angular 21 dashboard with FastAPI backend, PostgreSQL database, and PyPI package (portopt). Live at optimizer.silviobaratto.com.
Production-ready PPO implementation training AI to play Super Mario Bros. Features custom CNN architecture (4-layer with 512-unit FC), vectorized training with 8 parallel environments, and comprehensive evaluation tools. Achieves 80% success rate with 10M timesteps. Includes Docker support, TensorBoard/W&B integration, and modular codebase.
AI-powered dating simulation platform using LLMs to generate realistic profiles, simulate bilateral swipe decisions with compatibility scoring, and produce natural conversations. Multi-LLM support (OpenAI, Anthropic, Gemini, Ollama) with PDF report generation and analytics.
Benchmarking framework evaluating frontier LLMs (GPT-5.2, Claude Opus 4.6, Sonnet 4.6, Gemini 3.1 Pro) on numerical analysis problems. Validates LLM outputs against pre-computed expected values with automated pytest testing.
Probabilistic study of Italian Bingo (Tombola) using hypergeometric distributions and Monte Carlo methods. Analyzes winning probabilities across different game configurations.
PPO agent discovering optimal strategies for Cucu, a traditional Italian card game with incomplete information. Features curriculum learning, Monte Carlo baseline analysis, multi-agent self-play (4-8 players), and game visualization with animated GIF/MP4 rendering.
Lightweight text prediction system that learns from WhatsApp conversations to suggest words as you type. Pure Python implementation with zero dependencies, achieving ~100k words/second training speed. Features real-time terminal UI and privacy-focused local processing.
CLI tool for scaffolding full-stack projects with support for FastAPI or NestJS backends, Angular frontends, and Docker configuration. Optional authentication setup with token-based or Supabase auth.
Python machine learning library for stock market analysis implementing LSTM, BiLSTM, GRU, BiGRU, MLP, and Bayesian Neural Networks. Includes probabilistic models (NNHMM, DMM) for financial time series with scikit-learn-style interfaces. Built on PyTorch and Pyro with S&P 500 data downloader.
Conditional VAE generating handwritten digit images (0-9) from natural language prompts. Uses SentenceTransformer for semantic text encoding (384-dim) with 20-dimensional latent space. Demonstrates text-conditioned generative modeling fundamentals with KL annealing.
I'm always open to discussing new projects, creative ideas, or opportunities to be part of your vision.
Get In Touch