██╗███╗   ███╗   jmjm1558@github
     ██║████╗ ████║   🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧
     ██║██╔████╔██║   role ............... AI trainer · independent contractor
██   ██║██║╚██╔╝██║   location ..................... Colombia · remote (UTC-5)
╚█████╔╝██║ ╚═╝ ██║   focus ................. LLM agents that reach production
 ╚════╝ ╚═╝     ╚═╝   langs .................. Python · TypeScript · some Java
                      stack .............. FastAPI · Next.js · Docker · Gradio
                      models .............. Claude · DeepSeek · Whisper · Qwen
                      machine ............. Ubuntu 24.04 · RTX 5060 · 30GB RAM
                      server ............... Hetzner VPS · Tailscale · systemd
                      editor .............. VS Code + Claude Code · zsh · tmux
                      studying ............ numerical methods · modern physics
                      site .................................... juanjimenez.ai
                      🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧🟧

I build systems where the model does the work, not the demo. Most of what I ship runs on my own machine or a small VPS: local inference, agents wired to real tooling, and the unglamorous plumbing that keeps them alive at 3am.

+ available for AI training & evaluation work (remote, USD)
+ available for short ML / automation builds
- not taking full-time roles right now

agent-island · GNOME Shell extension

A Dynamic Island for your desktop. A pill replaces the clock in the top bar and shows every AI coding agent session running on the machine in real time, one dot each, colored by state: green means working, amber means it is waiting on you, gray means done. Click a row and it focuses that session's terminal pane. The card also mirrors MPRIS playback and the notification tray, so the notch becomes a hub instead of one more widget.

No daemon, no polling, no custom socket. It reads what the agents already write to disk. Disabling the extension puts your clock back exactly where it was.

GJS · GNOME Shell · tmux · MPRIS

transcriptor · local speech pipeline

Drop in a recording or paste a link, get a Markdown transcript with timestamps and an auto-written meeting summary. Whisper large-v3 runs on the local GPU, so the audio never leaves the machine and there is no per-minute bill.

The interesting part is the live-meeting mode: it records the mic and the speaker monitor as two separate tracks, then interleaves them by word-level timestamps into a labeled dialogue. A two-pass anti-echo filter throws away what the mic picked up from the speakers, so you get "me" and "them" instead of one soup of overlapping text.

Python · faster-whisper · CUDA · Gradio · yt-dlp · PipeWire

jimenez-site · personal site

Services, projects and contact in three languages. Locale routing (es/en/pt) runs through middleware with next-intl, so every page exists once and the copy lives in message catalogs. Adding a language is one JSON file and one line of config, not a new set of pages.

TypeScript · Next.js 16 · next-intl · Tailwind 4 · Framer Motion

• Tender-screening agent. Ingests Colombian public procurement notices, scores each one against a company's actual capabilities and emits a GO / NO-GO verdict with the risks spelled out, so a small firm stops reading hundreds of tenders it was never going to win.
• WhatsApp sales agent + CRM. Qualifies inbound leads in conversation and hands off to a human at the right moment, with the history already written to the CRM.

I keep a lessons file. Every entry is a day I am not getting back:

• A GPU dashboard that polls nvidia-smi keeps the dGPU out of runtime suspend permanently. It burned battery to measure how much battery was being burned. Reading sysfs instead does not wake the card.
• openwakeword's reset() only clears the prediction buffer. The ~1.3s feature window keeps its audio, so the assistant heard its own wake word from ten seconds ago and woke itself up, forever.
• My university's wifi sends a TLS reset to huggingface.co based on SNI alone. The code was fine. The Whisper weights now come from ModelScope.
• secrets.token_urlsafe() happily emits +, / and =. Put one in a redis://:pass@host:port URL and the parser tries to cast half your password to a port number.
• Models wrap JSON in a markdown code fence no matter how firmly the prompt forbids it. Strip the fence before json.loads. Every time. The smaller the model, the more certain it is that you wanted the pretty version.
• SECOP II has no public API and its documents sit behind a reCAPTCHA. The correct engineering answer was that this part does not get automated.

juanjimenez.ai · jmjm1558@gmail.com