Skip to content

Linux Infrastructure · Scientific Computing · Open Source

Dax

I run real Linux infrastructure — and build software for science.

I operate seven Linux servers — spread across my own hardware, my university and a rented VPS, and stitched into one private network with Headscale. Three of them run the scientific-computing pipelines we deploy for the CMS experiment at CERN, whose computer-science group I lead at Escuela Politécnica Nacional. I also keep a security platform running in production. One obsession runs through all of it: a robust, modern, secure architecture — in the code and in the infrastructure. I love science, I build in the open, and I would rather understand a system all the way down than ship something I cannot explain.

  • 7Linux servers operated
  • 86public open-source repos
  • 1,398contributions, past year
  • CMS · CERNscientific computing
See the workGet in touch

Work & research

  1. CMS Collaborator — computer-science lead

    CERN — CMS Experiment · through Escuela Politécnica Nacional

    May 2025 — present

    I lead the computer-science side of EPN’s group in the CMS collaboration, working on research reproducibility and on the computing that large-scale science actually runs on.

    • Reproduced an antimatter-search analysis with the Data Preservation and Open Access (DPOA) group — reproducibility, data integrity, and validation of the original result.
    • Deploy and benchmark Apache Airflow pipelines across three Linux servers for the BRIL group, supporting the CERN Phase-2 upgrades, and report the results in CERN meetings.
    • Designed and proposed modern, containerised architectures for scalable research pipelines.

  2. Research Assistant

    ADA Data Science Laboratory

    Mar 2025 — present

    Applied machine-learning research on optimization algorithms.

    • Researched optimization methods for deep neural networks, including second-order optimizers.
    • Designed and implemented the Python pipelines for training, evaluation and benchmarking.
    • First author of a comparative study on optimization paradigms in DNN training (presented; publication pending).

  3. AI Developer & Linux Systems Integrator

    WORB

    Apr 2025 — Dec 2025

    Full-stack development integrating AI into production workflows.

    • Backend architecture, development and integration.
    • Embedding search and LLM integration.
    • Database design, plus frontend development.

The CMS work is open — MiniParT — jet flavour taggingDark matter search — paper replicationCMS-EPN on GitHub.

Publications

  • Toward Computationally Efficient AI: A Comparative Study of Optimization Paradigms in DNN Training

    D. Navarrete, A. Flores-Reyes, A. E. Camino, G. Suntaxi, L. Recalde, D. Martinez-Mosquera

    Escuela Politécnica Nacional, Quito · Presented at conference · publication pending

    A comparative study of optimization paradigms for training deep neural networks — first-order against second-order and adaptive methods — asking not which one converges best, but what each one actually costs in compute to get there.

    First author. I designed the experimental methodology and ran the entire evaluation end to end: the training and benchmarking pipelines, every run and measurement, the figures, and the analysis the conclusions rest on.

GitHub activity

1,398 contributions in the last year, across 190 active days.

@daxrpm
Less
More

Selected work

Personal infrastructure

Infrastructure

Linux servers in three different places — my own hardware at home, machines at my university, a rented VPS — welded into a single private network with Headscale, so they behave like one machine no matter where they sit. Behind it I self-host the cloud I would otherwise rent: files, photos, media, monitoring. Reachable from anywhere I am, exposed to no one else.

Private infrastructure · the ground everything else of mine runs onLinuxHeadscaleDockerNextcloudImmichNginxTCP/IP

Fleet

Infrastructure

An agentless web console for a whole fleet of Linux servers — like Cockpit, but across every machine at once, with terminal, Docker, systemd, SFTP, a Headscale tailnet and Coolify built in. I wrote it because I wanted one secure place to operate the servers I already run.

Sourceupdated todayGoReactSSHHeadscaleDocker
Live CPU, memory, disk and network per server, collected entirely over SSH — no agent to install. Host IPs and OS versions are blurred on purpose.

Nexalink

Backend

A physical-security platform that runs in production at a private company: real-time access control over biometric hardware, role-based permissions, incident and ticketing workflows, guard operations and live tracking — on a hexagonal architecture designed to survive its own growth.

Private · running in production. Source and client withheld.LinuxHexagonal ArchitectureAccess ControlRBACReal-time Systems

Apogee

Scientific computing

A two-stage rocket ascent simulator — equations of motion, ODE integration, yaw-steering angles, full 3D — that uses the shooting method to optimize insertion into a circular Low Earth Orbit.

Source★ 2updated 5mo agoTypeScriptPythonThree.jsNumerical Methods
Full mission: liftoff from Pedernales, staging, and orbital insertion. The long coast through the dark is sped up; the launch runs close to real time.
Real elevation data — Quito at 2,850 m.
Orbit view after insertion.

dax-auth

Systems

A PAM module written from scratch in Rust that unlocks Linux with your face — on-device, encrypted at rest, with liveness detection. No cloud, no telemetry. People close to me log into their machines with it.

Source★ 1updated 1mo agoRustPAMLinuxComputer Vision
$ sudo dax-auth enroll
  → capturing 8 frames · encrypting to /etc/dax-auth
$ sudo -i
  dax-auth: face matched (0.31s) · auth sufficient
  root@fedora:~#

I/O & Filesystem Benchmarks

Systems

C benchmarks comparing O_DIRECT, sendfile() zero-copy, UNIX domain sockets and TCP sockets on ext4 — written to find out what the kernel really does, rather than take anyone’s word for it.

Sourceupdated 12mo agoCLinuxSyscallsSockets
$ ./bench --size 1G --fs ext4
  A  read/write + user buffer .... 1.94 GB/s
  B  O_DIRECT (no page cache) .... 1.21 GB/s
  C  sendfile() zero-copy ........ 3.78 GB/s
  D  UNIX domain socket .......... 2.40 GB/s
  E  TCP loopback ................ 1.07 GB/s

Dax Assistant

AI / ML

A self-hosted, voice-first AI assistant that runs locally, speaks to any LLM, and can act on my machine through a sandboxed tool layer — with every destructive call gated behind an explicit approval.

Sourceupdated 11d agoPythonFastAPIMCPReactOllama
Ask it anything, and it can act — here it reads what is playing, then books the calendar event.
Every destructive tool call is gated — the model proposes, you approve.
5 MCP servers, 204 tools, with an audit log of everything executed.
One provider port, six adapters — Ollama, Anthropic, OpenAI, Gemini, DeepSeek, Codex.

TexTeller-Rust

AI / ML

A native Rust port of TexTeller that turns a photo of a handwritten equation into LaTeX — byte-exact with the original Python implementation, and shipped as a single binary.

Sourceupdated 2mo agoRustONNX RuntimeMachine Learning
InputHandwritten equation: the square root of 2x over 2x equals the sum from i=0 to infinity of 3i + 2 + 5x
A photo of handwriting.
Output — LaTeX
\frac{ \sqrt{2x}}{2x}= \sum_{i=0}^{ \infty}3i+2+5x
Rendered
2x 2x = ∑ i=0 ∞ 3i+2+5x

EPN Tracker

Backend

A degree-progress and grade tracker for Escuela Politécnica Nacional students, with the academic rules of the university encoded as a pure domain core.

Sourceupdated 5d agoFastAPIPostgreSQLRedisReactDocker
The whole degree as a prerequisite graph — 50 courses, 135 credits. Tap one to light up its chain.
Answers the one question every student actually has: what do I need on the retake?
Weighted components, with eligibility computed from EPN’s real rules.

Drone Detector

AI / ML

Real-time drone detection with YOLO11 at ~11 ms/frame on a CPU — no GPU — with automated alerts straight from RTSP cameras.

Sourceupdated 3mo agoPythonYOLO11OpenVINOOpenCV
$ drone-detector --source rtsp://cam-01 --backend openvino
  YOLO11s · OpenVINO · 640×640
  [12:04:31] drone  conf=0.91  → WhatsApp alert sent
  [12:04:33] drone  conf=0.88  → cooldown (60s)
  avg 11.2 ms/frame · 89 fps · no GPU

Also on GitHub

This list updates itself from the GitHub API. See all 86 repositories.

Stack

Grouped, not rated. The work above is the evidence.

Infrastructure & DevOps

  • Linux
  • Docker
  • Headscale
  • Nginx
  • Kubernetes
  • Coolify
  • Apache Airflow
  • Proxmox

Networks & Security

  • TCP/IP
  • Network Engineering
  • Wireshark
  • SSH
  • RBAC
  • Keycloak
  • OAuth2

Systems programming

  • Rust
  • C
  • Go
  • PAM
  • Syscalls
  • Sockets

Scientific computing

  • Python
  • ROOT
  • NanoAOD
  • Numerical Methods
  • NumPy
  • SciPy
  • Pandas

Backend

  • FastAPI
  • PostgreSQL
  • Redis
  • Django
  • Spring Boot
  • REST
  • TypeScript

AI & machine learning

  • PyTorch
  • TensorFlow
  • scikit-learn
  • ONNX
  • OpenCV
  • MCP
  • Ollama

Most-used languages across public repos

  • Python15
  • Jupyter Notebook15
  • Java11
  • TypeScript8
  • JavaScript4
  • Shell3
  • Rust2
  • HTML1

Education & certifications

  • Escuela Politécnica Nacional logo

    BSc Computer Science

    Escuela Politécnica Nacional

    Nov 2023 — present

    Ecuador's leading engineering university. Algorithms, data structures, operating systems, numerical methods.

  • IBM logo

    AI Engineering Professional Certificate

    IBM

    2023

    Deep learning with Keras, TensorFlow and PyTorch; computer vision; neural networks.

    View certificate
  • IBM logo

    Data Science Professional Certificate

    IBM

    2023

    Python, SQL, data analysis and visualization, machine learning.

    View certificate
  • Escuela Politécnica Nacional logo

    Data Science Diploma

    Escuela Politécnica Nacional

    2023

    Statistical analysis, machine learning, data visualization. 160 hours.

    View certificate

About

I got into this from the bottom up — curiosity about how a machine actually works. That turned into a shell in C, then a PAM module in Rust, then benchmarking what the kernel really does when you ask it to copy a file without touching userspace. It is the same curiosity that makes me run my own servers rather than rent someone else’s abstraction.

Science is the reason for all of it. Reproducing an antimatter analysis with CMS taught me that a result nobody else can re-run is not a result — and the scientific-computing work I do now exists to make sure the pipeline behind such a result still runs on someone else’s machine, a year from now.

I am not interested in code that merely works. I care about the architecture underneath: clear boundaries, least privilege, encrypted at rest, no cloud when the device can do the job itself, and an audit trail for anything destructive. What I can open, I open — production systems stay private when a client’s security depends on it.

FAQ

Who is Dax Navarrete?
Dax Navarrete (GitHub: daxrpm) is a Computer Science undergraduate at Escuela Politécnica Nacional in Quito, Ecuador, and an open-source developer working across Linux infrastructure, networking, security and scientific computing. He operates a seven-server Linux estate, keeps a physical-security platform running in production, and collaborates on the CMS experiment at CERN through EPN’s group, which he leads on the computer-science side.
What does Dax Navarrete work on at CERN?
Dax Navarrete contributes to the CMS collaboration at CERN through Escuela Politécnica Nacional, where he leads the computer-science group. With the Data Preservation and Open Access (DPOA) group he reproduced an antimatter-search analysis, working on reproducibility and data integrity. He now works with the BRIL group on scientific computing: deploying and benchmarking Apache Airflow pipelines across a three-server Linux environment and reporting the results back to CERN.
What infrastructure does Dax Navarrete run?
Dax Navarrete operates seven Linux servers spread across his own hardware, his university and a rented VPS, joined into a single private mesh network with Headscale. Behind it he self-hosts his own cloud and media services — Nextcloud, Immich and others — and three of the servers run scientific-computing pipelines for the CMS group. He wrote Fleet, an agentless console in Go, to operate all of them over SSH from one place.
What programming languages does Dax Navarrete use?
Rust and C for systems programming, Go for infrastructure tooling, Python for scientific computing and machine learning, and TypeScript for interfaces. Notable work includes dax-auth (a PAM face-authentication module for Linux, written from scratch in Rust), Fleet (an agentless Linux fleet console in Go) and Apogee (an orbital ascent simulator in TypeScript and Python).
Is Dax Navarrete available for work?
Yes — open to internships, research collaborations and open-source work in infrastructure and DevOps, scientific computing, security and networking, or systems programming. Remote or relocation, internationally. Reach him at dax@daxrpm.dev.

Contact

Open to internships, research collaborations and open-source work — in infrastructure and DevOps, scientific computing, security and networking, or systems programming. Remote or relocation, internationally.

GitHubLinkedIn