/projects/archive
Exploration history.
Things I tried, built and ultimately stopped. Not failures — data points. Each one taught something real about products, markets and my own limits.
AI Routes Assistant
Build an AI assistant that generates personalized travel routes from natural language. Ask «give me 3 days in Porto for a coffee nerd» and get a structured, bookmarkable itinerary.
The LLM prompting for route generation was surprisingly good. Got to a working MVP fast.
No clear distribution channel. The product worked but getting users was a different problem I had not solved. Also maintaining accurate POI data was expensive.
Validation before building. Distribution is the product, especially for local-data dependent tools.
Amazon SEO with AI
Use LLMs to automatically optimize Amazon product listings — titles, bullets, descriptions — based on keyword research and competitor analysis.
GPT was decent at rewriting copy. The pipeline for fetching competitor data with scraping worked.
Amazon ToS makes aggressive scraping risky. The real constraint was API access to actual search volume data — everything was estimation.
Platform dependency is a real risk. If the platform controls the data you need, you are always one policy change away from dead.
AI Onboarding System
Replace static onboarding flows with a conversational AI that adapts the setup sequence based on user answers and behavior.
The conversational flow was engaging. Users stayed longer in the onboarding steps compared to static alternatives in early tests.
Integration complexity was high. Every client had a different stack. Turned into a consulting engagement rather than a repeatable product.
B2B tools need strong opinions about the stack or infinite customization budget. I had neither.
Invoice Generator with AI
Voice-to-invoice: describe what you did this week and get a structured, ready-to-send invoice in seconds.
The core flow worked. Speaking was faster than filling forms. The output was clean enough to send directly.
Too many free alternatives exist. Wave, Bonsai, FreshBooks. Hard to compete on price with existing solutions that do more.
The voice UX was a real insight — reused it in TellMe. Sometimes a failed product is a feature in another one.
MVP en 24 horas
Document the process of going from zero to a functional MVP in a single day — the decisions, the cuts, the order of operations.
The constraint forced clarity. Scope became obvious when you only have 24 hours. Great exercise for mental models around «what is an MVP really.»
Speed is a skill. The faster you can de-risk assumptions the better. This shaped how I approach every new project now.