Andre
A platform of AI agents that answer, qualify, book and sell over WhatsApp, with a built-in CRM, in production with real customers.
I make software that has to work first thing on Monday. I built the self-service system behind the Electoral Tribunal kiosks, where Panamanians identify themselves with their fingerprint and get a duplicate ID card or a certificate without queuing, and the queueing system behind the 2021 vaccination rollout. Now I am working on two products: Andre, an agent that answers a business’s WhatsApp, books appointments and hands the sales team leads that are already qualified, and El Empleado Juan, which tracks attendance with face recognition and gets payroll ready.
Now: Andre and El Empleado Juan.
A platform of AI agents that answer, qualify, book and sell over WhatsApp, with a built-in CRM, in production with real customers.
A B2B SaaS for biometric attendance and automated payroll for Latin American SMEs, with in-house facial recognition on low-cost hardware.
Parsing, expression trees, an inference engine and knowledge representation, with no library solving it underneath. A system that does not compute the answer: it reproduces the reasoning and explains it step by step.
One Second SolverAn end-to-end pipeline of my own: detection with MTCNN and RetinaFace, embeddings with FaceNet and ArcFace, anti-spoofing, transfer learning and edge deployment. On cheap hardware and integrated with third-party terminals — which is where most vision projects fall over.
El Empleado JuanSystems where being wrong about who someone is not an option: legally valid documents, payment taken on the unit itself, continuous operation and every transaction auditable. No "we will fix it next sprint".
Electoral TribunalSoftware governing people moving through a building in real time, across several sites at once, with a protocol to meet and no rehearsal. When somebody is waiting in front of it, availability is not a metric: it is the product.
National vaccinationMicroservice ecosystems with streaming and real-time processing: Node.js, PostgreSQL, MongoDB, Redis, Kafka, Spark and Flink on Kubernetes over AWS. Multi-tenant by design, which is the one thing you cannot retrofit later.
FluyappA Llama-style decoder-only model written from scratch: tokenisation, attention, training. It is what lets you know why the model does what it does, and above all what its architecture cannot give you no matter the prompt.
WritingRetrieval, embeddings, tool use and integration with real systems, evaluated against questions with known answers. Agents that query live data and have an explicit path to saying "I do not know" instead of making it up.
RAG agentsAgents that qualify, book and move the pipeline on real systems, with the AI proposing and the system validating: typed schema, the user’s own permissions and constraints in the database. Stateful flows on LangGraph, multi-tenant with Row-Level Security and the official WhatsApp API.
AndreFrom a WhatsApp bot to a multi-tenant platform with three capture layers, statutory payroll with auditable premiums, feature flags, licensing and CI/CD. A product with paying customers, not a prototype.
El Empleado JuanI am Colombian and live in Panama. I started at university with a symbolic language app that worked like a teacher, was CTO at XPLOR, co-founded Fluyapp and El Empleado Juan, and now I am working on Andre.
As a kid I took everything apart to see how it worked. Still do, except now it is AI models.
Universitat Rovira i Virgili, Spain. Before that, Computer Systems Engineering at the Technological University of Panama.
I help physicians in Panama with the statistics in their theses: the analysis design, the models and the figures they have to defend.
Work in the public sector? I have a concrete proposal, with two government systems already running behind it.
Read the proposalIf you are thinking about building something with AI and want to talk it through with someone who has put it into production, write to me. I answer myself.
hola@giron.aiUsually the same day.