Juan Saint Amant saintamant

Juan Saint Amant

Data Engineer — AI Systems, Agent Orchestration & Automation
Buenos Aires, Argentina

About

Data Engineer focused on AI systems, multi-agent orchestration, and production-grade automation. I design and build end-to-end solutions — from OCR pipelines and credit scoring models to trading bots and AI command centers with 18 orchestrated agents.

Currently building Saint Tower: an interconnected AI platform where agents communicate, transfer context, and execute workflows autonomously — managed via Telegram from a pixel art command center.

My work spans:

AI Agents & Orchestration — Multi-agent systems with inter-agent messaging, context transfer, automated workflows, and Telegram-based remote management
Machine Learning — Credit scoring models, trading strategies (Safe Yield, Arbitrage, Whale Shadow), XGBoost, scikit-learn
NLP & Document AI — RAG systems with specialized agents, natural language-to-SQL for 305+ table schemas, OCR with GPU acceleration (100% accuracy)
Full Stack — Next.js + React frontends, Python/FastAPI backends, real-time systems with WebSockets, PixiJS pixel art rendering
Infrastructure — Self-hosted Linux server running 7 AI services on a single RTX 3050: Docker, Ollama, Airflow, PostgreSQL, Home Assistant

Featured Projects

Saint Tower

AI command center orchestrating 18 agents across 8 offices with a pixel art interface (SNES style). Each agent is a real Claude Code session with its own repo, context, and tasks. Agents communicate via animated envelope messages, run automated morning workflows (KPI analysis, lead research, metric processing), and are managed remotely via OpenClaw — a Telegram bot that interprets natural language intent and dispatches tasks through Saint Tower's API.

Next.js PixiJS WebSockets SQLite xterm node-pty Claude

OCR Pipeline

Production OCR system extracting salary data from payslips with 100% accuracy. 4-tier processing pipeline with GPU acceleration: raw extraction → field detection → validation → structured output. Handles multiple document formats, noisy scans, and edge cases through cascading confidence checks.

PaddleOCR CUDA Python RTX 3050

RAG Collections System

AI collections assistant with 4 specialized agents trained on business methodology. Each agent handles a different stage of the debt collection process. Built with vector embeddings for methodology retrieval and local LLM inference for privacy-sensitive financial data.

ChromaDB Ollama FastAPI Python RAG

SQL AI Agent

Natural language to SQL interface for schemas with 305+ tables. Production deployment at a financial institution. Handles schema disambiguation, multi-table joins, and query validation. Users ask questions in plain Spanish and get results in seconds.

LLMs Python SQL Server Claude

Credit Scoring Model

ML credit risk model running in production at a real fintech. Feature engineering from messy Argentine financial data, model training with cross-validation, and a visual dashboard for risk analysts. Processes loan applications and outputs default probability scores.

scikit-learn XGBoost Python Power BI

Polymarket Trading Bot

Automated prediction market bot with 3 ML strategies: Safe Yield (>97% probability outcomes for 1-3% per cycle), Arbitrage (YES + NO < $1.00 detection), and Whale Shadow (successful trader signal confirmation). Includes Telegram alerts for trade execution and P&L tracking.

Python scikit-learn XGBoost Polymarket API Telegram

Meeting Scribe

Meeting transcription and summarization tool. Records audio with device management, transcribes with segment merging, and generates AI-powered summaries with custom prompts. Persistent storage for searchable meeting history.

Python LLMs Audio Processing

Health Tracker

Health tracking system integrating Apple Watch data with a Telegram bot. Auto-detects workouts via Apple Health export, tracks daily weight, logs food habits, and visualizes everything in a web dashboard. Airflow sends automated alerts if no workout is detected by end of day.

Python Flask Gemini Airflow SQL Server Chart.js

AI Study Platform

Certification prep system with 12 specialized tutor agents for the NVIDIA GenAI LLM certification. Features a progress dashboard across 72 topics and 5 weighted domains, mock exam simulator (50 questions, 60 minutes), flashcards, and an AI professor that adapts to your progress in real time. Built in 1 hour.

Python Jupyter Claude Vanilla JS

Finance Dashboard

Personal finance tracker with expense categorization, income tracking, subscription management, savings goals, installment tracking, and USD/ARS currency monitoring. Waterfall charts for monthly closures and trend analysis across months.

Next.js React Supabase TypeScript Tailwind

Most projects are in private repos. Saint Tower (open source) is available for exploration.

Self-Hosted Infrastructure

Everything runs on my own hardware — no cloud dependency:

🖥️  Linux Server (Pop!_OS 24.04)
├── 🧠 Ollama — local LLM inference
├── 🔄 Apache Airflow — workflow orchestration
├── 🐳 Docker — containerized services
├── 🏠 Home Assistant — home automation
├── 🐘 PostgreSQL — databases
├── 📊 Portainer — Docker management UI
└── 🎮 GPU: RTX 3050 — 7 AI services on a single card