Luis Embon Strizzi

Listings optimized with AI and real marketplace data: Prompty scans a listing, audits it against the top sellers of its category, and generates optimized titles, descriptions and attributes in seconds. I own the backend: FastAPI, DSPy pipelines compiled with MIPROv2 on Claude, Supabase and the Mercado Libre OAuth integration. Built with Ciro Vilmer & team — now in public beta.

An end-to-end agentic platform for fintech disputes: a Claude agent with a real tool-use loop identifies the transaction, requests evidence and opens the case; RAG with citations grounds every resolution in internal policies, central-bank regulation and card-network chargeback rules; n8n orchestrates SLAs, notifications and escalations; and every agent decision is traced for auditability. FastAPI, Postgres + pgvector, MongoDB and Next.js — the whole stack boots with one docker compose up.

In a roundabout the same past can continue toward several exits, so single-trajectory models average modes into physically impossible paths. We modeled p(y|c) with a Conditional VAE and generated K=20 plausible futures, climbing a 13-model ladder from constant-velocity baselines to a TCN-encoder CVAE, reaching minADE₂₀ 0.517 m and 0.991 exit-intent accuracy, plus a 2D latent that reads as a maneuver selector. Built as my ML course capstone at UdeSA and presented as a poster at AI Fest 2026 — try the live demo: the model prunes 20 futures down to one as evidence arrives.

I authored the Artificial Intelligence annex of a pediatric dermatology textbook (Editorial Panamericana), written with a physician from Hospital Italiano de Buenos Aires: fundamentals, applications and limits of AI in the specialty. The chapter is backed by a working prototype — a transfer-learning classifier over ~40k dermatology images across 10 conditions, with Grad-CAM visual explanations behind the book's figures.