AI experiments by Irina Vélez — from first API call to robots and blockchain agents
experiments logged.
The ones that shipped. The ones that didn't.
All of them.
A research paper near the North Pole. ML models in R. The first encounter with what AI could become.
2 experiments logged
Research paper exploring AI as a general-purpose technology driving economic growth, with focus on intangible assets in economic transition.
This paper opened my eyes to AI. At Hanken, near the North Pole, I gained access to ArXiv, discovered Hinton, Brynjolfsson, the Stanford AI Index — and realized I was standing in front of something different.
Machine learning models to predict poverty classification in Colombia using national survey data.
First API calls. First RAG. First app. Everything broke — but things kept working.
4 experiments logged
My first RAG assistant — MongoDB vector search over 19 World Economic Forum reports. I didn't know what I was doing, but it worked.
The answers were accurate. I fed it WEF reports and it could synthesize insights across documents. This was my first 'it actually works' moment with AI. The repo is gone, but the memory isn't.
Baby steps with the OpenAI API. My first scripts calling an LLM programmatically.
First app built with Cursor — consumed the Spotify API. My introduction to backends, APIs, and how apps actually work.
AI-powered news aggregator — fetches from TechCrunch, MIT Tech Review, Ars Technica, summarizes with BART, extracts insights with Gemma.
Multi-agent changed everything. Started building for others — live, every Friday. First real client.
14 experiments logged
Enhanced RAG app combining PDF analysis with real-time web search. My most starred project.
RAG had evolved massively since Ari. This time I understood what I was building. Combining document knowledge with live web data was the unlock.
I didn't want to invent problems that didn't exist. So I asked LinkedIn what people needed and built solutions live, every Friday.
Building for others teaches you faster than building for yourself. The 7 experiments that came from this proved I could ship, even when I didn't have a product of my own yet.
5 experiments born from this
Feedback questionnaire app to understand tech debt from companies. Built on Replit, struggling to make a simple form work.
AI research platform for IT professionals — 5 specialized assistant modes powered by Perplexity AI, focused on telco infrastructure.
My first AI course — interactive HTML apps explaining fundamental AI concepts, entirely in Spanish.
First multi-agent system — a Researcher agent (Perplexity) gathers info, an Analyst agent (GPT-4) synthesizes professional reports.
Learning Claude Code by attempting to clone Lovable — real-time streaming code generation with Docker.
n8n workflow to curate AI news from Reddit and HN. My frustrated attempt at building a news curation system.
WhatsApp bot for building maintenance — my first project with a real SMB client. Classifies requests, generates work orders, dashboard.
SF. Hackathons every weekend. Robots, voice AI, blockchain agents. What takes months elsewhere takes days here.
3 experiments logged
Multi-agent geoscience tool — two AI expert agents (geotechnical + AI) collaborated to build this borehole visualization app. I didn't vibe-code it, the agents did.
3 hands-on AI attack vector workshops: prompt injection on resume screening, memory poisoning via vector DB, MCP tool description poisoning.
Solana hackathon project — 5 AI agents buy and sell skills via Metaplex Agent Registry with x402 payments. Agent-as-a-Service economy.
Physical AI concierge robot — 3-wheel robot with iPhone, ElevenLabs voice, Claude brain.
Built at a hackathon in 24 hours. Hardware + voice AI + LLM in one weekend.
Pitch your startup in 60s and get roasted by AI personas of PG, Ben Horowitz, and Marc Andreessen.
Gemini Live API enables speech-to-speech AI that feels human. Voice is the next frontier.
Daily AI intelligence briefing for LATAM professionals.
From a frustrated n8n workflow to a live product people visit daily.