about
Aroa Xinping
Aroa Index documents a model. This documents the person who trained it — in the same format, because it still feels like the one that fits best.
class Aroa:
"""Perfil · aroaxinping · 2026-08
Rol: analista de datos en formación.
Origen: Bellas Artes → FP → Data Science (UOC).
Uso previsto: análisis, visualización y contarlo.
Training data: ver abajo.
Caveats: sesgada hacia lo que me interesa a mí.
"""[a] training data
my story
Data analyst in training. I come from design: I started in Fine Arts and got to data for the same reason I started drawing — to understand how things work so I could then explain them.
I'm doing the Data Science degree at the UOC and the Ironhack Data Analytics bootcamp at the same time. I mostly work with Python (pandas, NumPy), SQL and visualisation tools.
My first real project was my own account. I make content about data and technology on TikTok and Instagram, and I analyse my own metrics to understand what works and how an audience behaves — the project where I practise everything I study, with real data and real consequences.
Design didn't go anywhere: it's still here, mixed in with the data. An analysis nobody understands isn't finished, and that's exactly where the two things meet.
$ git log --graph aroa/data-journey
* a1f3c9e (HEAD → main) 2026 right now ● in progress
+ Data Science degree (UOC)
+ Data Analytics bootcamp (Ironhack)
+ Aroa Index: trained my own model for this site
+ Inferential statistics in progress...
* d8b02a1 2025 building foundations ✓ merged
+ Python: pandas, NumPy, Matplotlib
+ SQL from zero to advanced queries
+ First real project: my own metrics
+ Git and version control
* 3e91fc7 2024 the beginning ✓ merged
+ Design fundamentals (Fine Arts)
+ First visualisations and dashboards
+ Found out that data tells stories
* ??????? (origin/feature/ml) next coming up ⋯ pending
? Machine learning beyond TF-IDF
? Cloned voice across the whole model, not just index
# this branch doesn't exist yet. soon.
stack
What I use, by how much I use it
No percentage bars: nobody believes them and I would not know how to defend a “Python 90%”. It is grouped by something you can actually check — how often I really touch each thing.
software
- every day
- python · pandas · jupyter · git What I open without thinking.
- in projects
- sql · numpy · scikit-learn · matplotlib · seaborn · plotly · streamlit · prefect · aws s3 · astro I have used it end to end at least once.
- borrowed syntax
- rust · c++ · shell I understand the problem and I know how it is solved; what I lack is the language. Here AI gives me the syntax, not the approach.
running locally
- runner
- ollama
- general use
- llama3.1:8b · qwen3.5
- reasoning
- deepseek-r1:8b
- the one that does not fit
- gemma4:26b
A 26B at 4 bits needs about 15 GB and the card has 8, so ollama splits it and sends the rest to RAM. It runs, slowly. That is why the other three are 8B.
hardware
- work
- MacBook Air M1
- training locally
- PC propio · RTX 3060 Ti 8 GB · 24 GB RAM
- filming
- DJI Osmo Pocket 3
- typing
- Logitech Pebble
- listening
- Sony XM6 · AirPods Pro 2
What I keep about you
Nothing that identifies you. No cookies, no analytics, no browser fingerprint, no session: if you come back tomorrow, this site does not know you had been here.
The only thing stored are the questions the model could NOT answer — the text, its score and the language, nothing else. The ones it does answer are not stored.
There is a specific reason: the conclusion of this whole project is that growing the corpus pays off more than changing the algorithm, and without knowing what people ask there is no way to decide what to grow. It is the only piece of server here, and the code is in the open like everything else.