A small, honest portfolio of AI systems and workflows. Each entry describes the useful part, the boundary around it, and what the record does not prove.
A local-only service that accepts documents from a configured WhatsApp contact, runs an existing processing workflow, and returns the result to the same chat. The implementation keeps the webhook on localhost and pairing data on the server.
An evidence-backed research and lead-intelligence system with mission planning, public-source crawling, confidence checks, durable exports, and a private knowledge service. It is designed to preserve the source trail behind an answer.
A local pipeline for transcription, concept analysis, original scripting, Romanian voice, generated visuals, deterministic video assembly, and technical QA. The recorded run completed preparation, but publishing remained explicitly disabled and human-controlled.
I would rather show the boundary than hide it. A useful AI build is not just a demo that works once. It needs a clear input, a repeatable workflow, inspectable outputs, and an honest account of where a person still needs to make the call.