Rodrigo Medrano

Infrastructure Monitoring

Jan 2026 - Jan 2026 • 0 mos

The system ingests metrics (CPU, memory, latency), processes them in the background, and triggers email alerts when thresholds are exceeded. All of this happens in real time: a WebSocket broadcasts every metric and alert to the dashboard the moment it occurs. While it may look simple on the surface, the internal implementation is robust: - An asynchronous worker that persists data to SQLite in batches (10,000 metrics per request, returning a 202 or 429 status code if the queue is full). - A cooldown mechanism to prevent duplicate alerts, and default rules that do not overwrite existing ones. - A composite index ensuring that an old spike does not re-trigger an alert in the absence of new data. The dashboard uses vanilla HTML/JS with Chart.js, served directly by the FastAPI backend. Stack: FastAPI + asyncio, SQLAlchemy 2.0 + aiosqlite, Pydantic v2, WebSockets, SMTP, Docker.

Assistant SaaS

Oct 2025 - Nov 2025 • 1 mo

An AI assistant that independently reads, categorizes, and responds to a company's emails. I built it as a multi-tenant SaaS: multiple companies use the same instance without data mixing. When an email arrives, a LangGraph-powered agent classifies it (inquiry, appointment, refund, or complaint), retrieves context from a knowledge base using RAG, and drafts a professional response. The most interesting part: not everything is sent automatically. Sensitive cases (refunds, complaints) are held until a human approves them via a web dashboard. The AI ​​assists, but the final decision is human. Technical details: - RAG using pgvector + local embeddings (BGE), avoiding reliance on external vector services - Celery + Redis for asynchronous processing - Real integrations with Gmail and Google Calendar (plus a mock mode for development) - Tokens encrypted with Fernet (AES-256) Stack: FastAPI, PostgreSQL + pgvector, LangGraph, Gemini, Celery, HTMX, Docker.

Expense Analyser

Jun 2025 - Jun 2025 • 0 mos

An AI-powered expense analyzer: upload a PDF invoice or a photo, and the app reads it, classifies each line item into its accounting category, and displays the data on a dashboard featuring charts and budget alerts. Under the hood: - Data extraction using vision and text models via Groq (falls back to vision-only mode if the PDF lacks extractable text) - Batch accounting classification to minimize token usage, with JSON output - Financial precision using the Decimal type (never float) and ROUND_HALF_UP rounding - Accounting immutability: once accepted, an invoice cannot be edited or deleted - Duplicate prevention: each file is hashed using SHA-256 - SQLite for development, PostgreSQL for production; all managed via Docker Compose Stack: Streamlit + Plotly, Groq (LLMs), openpyxl, PostgreSQL, Docker.