I build AI products that solve real problems
From AI-powered financial analysis to real-time language learning: three shipped products with real users and real payments, built end to end.
Projects
AI applications built from problem to production.
Spralingua
Real-time AI voice tutor — speak into your browser and a streaming agent replies in under 2.5 seconds, then scores both your lesson goals and your pronunciation.
TubeText
AI-powered YouTube transcript extraction with LLM summaries, real-time streaming translation, and audio transcription — deployed as a production SaaS.
AgoraFinancials
Multi-LLM stock evaluation engine that orchestrates 9 AI models to analyze SEC filings, debate disagreements, and produce bias-reduced financial ratings.
The Problem
Speaking a new language out loud in real time is the skill apps skip. Vocabulary drills don't prepare you for the moment someone is waiting for your reply — and they never tell you whether you were even understandable.
The Approach
A real-time voice agent: a per-client Pipecat pipeline streams browser audio through Deepgram speech recognition, a Cerebras LLM on LangGraph, and MiniMax voices over a WebSocket. Silero VAD gates each turn so replies never fragment.
The Result
Live at spralingua.com with sub-2.5-second turn latency and true multi-user concurrency. Every session is scored on two axes — goal completion via an LLM judge and pronunciation via Azure Speech — across 7 locales and CEFR A1–B1 lessons.
Key Decisions
Per-client pipelines for structural multi-user isolation. VAD-gated buffering to stop fragmented replies. Evaluation runs post-session so scoring never adds to live latency. Lessons are YAML files, not code.
The Problem
YouTube holds massive knowledge but extracting insights from long videos is inefficient. No tool combined transcript extraction, AI summarization, and multilingual translation in one pipeline.
The Approach
Async FastAPI backend orchestrating multiple AI services: Deepgram Nova-3 for audio-to-text, LangChain + OpenAI GPT-4 mini for structured summaries, Cerebras Llama-3.3-70b for real-time streaming translation via SSE. PostgreSQL with async SQLAlchemy and Alembic migrations. Dockerized and deployed on Railway.
The Result
Production app at tubetext.app. Transcripts in seconds, AI-generated summaries, streamed translations (Spanish, French, German, Portuguese), and PDF export. Google OAuth + Stripe integration for billing.
Key Decisions
FastAPI for async-first performance. LangChain agents with YAML-driven prompts for maintainable AI pipelines. SSE streaming for real-time translation feedback. Deepgram over Whisper for production reliability. Docker + Railway for reproducible deployment.
The Problem
Stock analysis is drowning in information. LLMs can condense SEC filings into narratives, but a single model can hallucinate numbers, cherry-pick data, and bias its conclusions. No existing tool cross-validates AI-generated financial analysis.
The Approach
Built the Pythagoras Method: multiple LLMs independently evaluate 8 financial metrics, their ratings are harmonized, and conflicts trigger structured multi-round debates until consensus. Two-tier model architecture (fast for numbers, deep for RAG-powered qualitative reasoning) with OpenRouter as unified gateway.
The Result
Production platform at agorafinancials.com. 9 LLMs across 5 providers running in parallel, SEC filing RAG with pgvector, PDF report generation with scoring, Stripe billing with token economy, and watchlist with live scores.
Key Decisions
OpenRouter for unified multi-provider billing. LangChain create_agent with retrieval tools for deep-tier models. pgvector over Pinecone for simpler single-database infra. Sliding-window debate compression to keep token costs flat across rounds.
I'm an AI Application Engineer based in Germany. Originally from Mexico, I made the move, picked up German, and grew into software the long way around — close to engineering teams first, then building alongside them.
Now I ship AI applications end to end. Besides the deployed products shown above — each with real users and real payments — I also take on client work in voice-cloning and security-first internal tools. My stack runs from LangChain agents and RAG pipelines to Stripe billing and Docker deployment.