bruno arnaiz

Agentic OS developer

Bruno Arnaiz Santamaría

Advanced AI user · building agentic operating systems already in use at small businesses.

Based in
Madrid, Spain
Studying
Data Science & Engineering · URJC
Status
Looking for technical founders
00 /

Now

Updated Nov 2026

Building

Focused on two projects that are identical under the hood and very different on the outside: Neuvora, for businesses, and StudyFlow, for students. Same stack, same architecture — different customer.

Exploring

The problem I keep coming back to: making AI agents — whose outputs are inherently unpredictable — coexist with the traceability and rigor that a database, and every channel connected to it, demand. Around that core: fine-tuning open-source models, running agents inside Supabase Edge Functions, and strategies that let an AOS grow without turning into a chaotic pile of data.

Where to find me

I’d like to go to more events for founders and developers, where I can meet people who share my ambitions and interests.

Know a good one? Tell me about it.

01 /

Ventures

What I run. Two ventures, one architecture. Both private — happy to show them live.

Neuvora AOS: business departments feed one database; agents read and write to it; a control panel with human review on top. acquisition sales operations support database a1 a2 a3 agents control panel human review illustrative schematic

Neuvora AOS

Agentic operating systems for real businesses.

I find businesses and analyze them as a full loop, end to end — from acquiring a customer to selling the product. I automate as much as possible, down to a few human supervision steps, and deliver the result as an application.

My part
I start by mapping the business into departments and design the database around where each agent will read and write. Then I build the agents one by one, with special care for the traceability of every run and each agent’s guardrails.
Stack
Supabase (Postgres + Edge Functions) · Vercel (deploy + AI Gateway) · Docker on Debian · Python · TypeScript
Result
In production: a private civil-service exam syllabus platform, custom-built for a client.
Status
Systems in production, more in development.
StudyFlow: course material goes into one database; LITS, the tutor agent, reads it and answers through several channels. notes syllabi exercises database LITS tutor web Telegram notebook illustrative schematic

StudyFlow

The university virtual classroom of the future.

A virtual classroom with AI built in natively. It goes deep on many fronts: handwriting, note-taking, exercise generation, precise progress tracking — and LITS, an agent that works as a personal tutor, reachable from several channels with instant access to all course material: notes, syllabi and exercises.

My part
All of it, end to end.
Stack
Same stack as Neuvora · Supabase (Postgres + Edge Functions) · Vercel · Telegram
Result
Business plan done. Goal: a pilot running at a university by September 2027.
Status
In development, with real test users (free of charge) who help me polish it.
Link
No public link yet.
02 /

Open source

Small public repos: clone them, read the code, run them.

Terminal output of agent-regression on Claude: with prompt v2 the order-lookup case regresses from 2/2 to 0/2 because the lookup_order tool is no longer called.
A real run on Claude: a prompt tweak meant to cut costs made the agent stop looking up orders. Only the tool check caught it.

agent-regression

Regression tests for AI agents.

Change a prompt, a model or a tool and find out what you broke before your users do. Test cases live in a YAML file; every run is compared against a saved baseline, so you see what regressed and what got fixed — with an exit code your CI understands. It checks what the agent does (the tools it calls), not only what it says.

Stack
Python · YAML · pytest · GitHub Actions · Claude API
Status
Public · MIT license · tests passing on Python 3.10–3.13
03 /

How I build

If you’ve made it this far, something caught your eye. Here’s the big picture: how an agentic operating system really works when it’s designed from day one to integrate AI natively.

First, you lay out on paper all the information you’ll handle — down to the least significant field. Separately, you list every process that people or programs do today and that AI can genuinely take over. With that, you can build the database tables and the agents that read from and write to them. That way, no matter how much the database grows or how much time passes, the agents leave a trace of their work, and the AI can understand where the business is heading — its strengths and its weak spots.

That’s why Neuvora and StudyFlow are technically the same thing: same stack, same structure, different customer. Neuvora is for business owners; StudyFlow is for students.

agents ⇄ database ⇄ control panel

a well-structured database agents that read & write to it rules & traceability a real AI employee

Three rules I build by

  1. 01

    Plan first, code second.

  2. 02

    No agent ships to production until the traceability of its runs is verified.

  3. 03

    Commit, commit, commit — never put off till tomorrow what you can commit today.

04 /

Stack

What I actually use, not everything I’ve touched.

Languages

Python (local agents) · TypeScript / JavaScript (Edge Functions & web) · SQL · R

AI & agents

Closed models via Vercel AI Gateway · open-source models with Ollama (Llama 3 8B, Qwen 8B) · fine-tuning open-source models · multi-agent orchestration · tool calling & MCP · guardrails & run traceability

How I work

Claude Code (daily driver) · Codex · Kimi Code · Hermes · OpenClaw · my own agents

Data

PostgreSQL · Supabase · pandas · NumPy · R + Quarto · regression models · process optimization

Infra & deploy

Supabase Edge Functions · Vercel · Docker · Debian / Linux · systemd · Railway · Git & GitHub

Automation & channels

n8n · Telegram bots · scraping · voice with Twilio, Deepgram & ElevenLabs

Web

Next.js · React · HTML, CSS & JavaScript

05 /

About

I’m based in Madrid and study Data Science & Engineering at Universidad Rey Juan Carlos. I’m ambitious and hungry for new professional challenges.

I’m looking for people who want to build and sell — people ready to team up and go all in until the goals are met.

Beyond the technical side, I’d describe myself as honest, loyal to my friends, and serious — but fun when the moment calls for it.

06 /

Contact

If my work caught your interest and you have a technical profile with something to bring, write to me. I read every message.

I also take on quick, focused freelance work: small businesses or one well-defined idea.