reading level: variable, on purpose

Paulina Valenzuela
Lagos(N.(n.)S.s.)

范寶琳(S.N-)

Solo engineer, computational linguist, Spanish instructor. I build the tools that decide how hard a text feels, and teach the humans reading it. No team, no lab budget: everything on this page was designed, coded, tested, and shipped by one person. Me.

Two ways to read this. is a 3-minute skim of what I do; takes about 20 minutes and covers why. Switch anytime from the top-right — or the ☰ menu on mobile.

Where I'm headed: EdTech and health tech — language-learning tools that adapt to the reader, and patient-facing tools that make hard-to-see conditions legible. Academia stays firmly on the table; the teacher part never leaves.

Field: NLP × applied linguistics × language pedagogy Base: Taipei, Taiwan Status: Ph.D. candidate, Digital Learning & Education, NTUST

A1 origins

Where this started

First-generation college graduate. B.A. in Letters (linguistics & Hispanic literature), Pontifical Catholic University of Chile — first published paper in 2019. B.Ed. in Secondary Education on a government Teacher's Vocation Scholarship, plus a year teaching in the U.S. (Union College, NY).

I'm part of the first generation in my family to go to university (except for my aunt Nery). With a lot of interest, and no small amount of difficulty, I earned a Licenciate in Letters, with a mention in linguistics and Hispanic literature, at the Pontifical Catholic University of Chile (PUC). My parents always said education was the one inheritance they could give me. That stuck.

I'd always been interested in Japanese, Chinese, and Korean culture, so I took the chance to minor in Asian Studies and join the Study Club Asia. That's where I made my first close friendships with people from Hong Kong, Taipei, Tokyo, Incheon, and elsewhere.

Making the most of those opportunities, I co-organized an international seminar on Korean studies, placed 4th in a national essay on Korea, and (funding the trip myself with a part-time job) spent a summer studying Korean at Seoul National University.

Learning Korean was deeply satisfying and opened up new interests and dreams for me. Sadly, I haven't been able to keep advancing my Korean since, but it's a dream I haven't given up on.

My undergraduate thesis grew out of a small moment: my Asian friends became my first students. And among the many questions they asked, like why Spanish articles (el, la, un, una…) were so hard to get right, I often didn't have a satisfying answer. So I researched it. The result, "Uso y adquisición de artículos en español como segunda lengua," was published in Logos: Revista de Lingüística, Filosofía y Literatura in 2019, co-written with my advisor, Dra. Gloria Toledo. I finished the underlying research while recovering from a car accident, which taught me early on that research is one of the few things that travels well through a hard year.

A second degree, in Educación Media (secondary education), came with a Teacher's Vocation Scholarship from the Chilean government, while I also completed a Diploma in Teaching Spanish as a Foreign Language on a scholarship as the diploma's teaching assistant. After that, a year at Union College in New York as a language assistant was where I first started experimenting with technology (VR included) in the language classroom.

That teacher's-vocation scholarship didn't end at graduation — it turned into eighteen classrooms across three countries.

A2 teaching

One classroom at a time

Before "researcher" was on the business card, "teacher" was — and still is. Spanish, English, kids, adults, DELE candidates, exchange students, mining engineers learning survival Spanish: I've taught most of it, across three countries.

18 teaching roles across Chile, the U.S., and Taiwan (2015–present): K-12, adult education, DELE exam prep, corporate language training, and an EMI Teaching Assistantship at NTUST (Best Tutor award, 2021).

K-12Adult EducationDELE Prep EMICurriculum DesignEdTech (VR / ChatGPT) Corporate Language Training
2015Middlebury College Chile — Spanish Teacher (Santiago)
2016Intercoined Chile — Spanish Teacher (Santiago)
2016Colegio Los Nogales — Language & Communication, 11th grade (Santiago)
2017CEE El Golf — Academic Assistant & Spanish Teacher, SIELE support (Santiago)
2017Bridge Language Group — Spanish Teacher (Santiago)
2017University of the Andes — Spanish Teacher, exchange students (Santiago)
2017–18Union College — Spanish Language Assistant (Schenectady, NY)
2018Liceo Carmela Carvajal de Prat — Spanish Teacher, 7th–10th grade (Santiago)
2018PUC — Spanish Teacher, Haitian immigrant program (Santiago)
2018–19CEE El Golf — Spanish Teacher, Japanese mining-sector learners (Santiago)
2019Saint George College — Spanish Teacher, 10th–12th grade (Santiago)
2019–20MingZhi Junior High School — English Tutor (Taipei)
2019–21Freelance — Private Spanish Teacher (Taipei)
2021NTUST — EMI Teaching Assistant · won Best Tutor award
2021–24Lion and Lion — Spanish & English Teacher, DELE B1/B2 prep (Taipei)
2023Lifelong Learning Center — English Storyteller (New Taipei)
2023–Gram Cram School — English Teacher, grades 1–6 (New Taipei)
2024–26NTUST — Teaching Assistant, Applied Foreign Languages; helped bring ChatGPT into a Reading & Writing classroom

scroll for the full ledger — it's long on purpose

Off the page

Teaching was never only classrooms — though it was often those, in Chile and in Taiwan. It was also a post office in Santiago, where Japanese engineers hunted vocabulary through the stamp displays, and, for a few years, a soccer pitch and a basketball court.

Children’s faces are blurred on purpose. Consent matters more than a good photograph.

reading

What I'm reading and writing

A quick detour before the research gets technical: these are my reading profiles, on tools built by independent creators, not Amazon or Google. No sponsorship here, check them out if you're curious, it's genuine interest.

Substack newsletter ↗

Thousands of students later, one question kept resurfacing: why do some texts feel so much harder than others? That question became a dissertation.

B1 research

What "readability" means in my hands

Ph.D. candidate, Digital Learning & Education, NTUST. Dissertation builds a Spanish readability model (238 linguistic features, ~0.96 QWK, ~87% accuracy on 822 CEFR-labelled texts) and validates it against real learner perception. M.A. thesis (2020–2022) compared graded readers vs. authentic texts. Presented at 4 international venues; 1 publication, 1 in review, 2 more queued.

NLPReadability ModelingTransformers (BERT / BERTIN) Feature EngineeringCatBoostCorpus Linguistics Statistical ModelingMixed-Methods ResearchPythonR

My doctoral dissertation, "Assessing Text Readability in Spanish as a Foreign Language: Computational Modelling and Human Judgment," is a two-study pipeline: build a model that grades Spanish text difficulty as reliably as a trained rater, then check whether real learners actually experience difficulty the way the model predicts.

Study 1 — a hybrid readability model

A specialized, CEFR-labelled corpus (graded readers and DELE exam passages) feeds a pipeline of 238 linguistic features across six families — lexical, syntactic, cohesion, traditional, morphological, vocabulary — extracted in part with three original tools, SpanishMorphAnalyzer, SpanishComplexityAnalyzer, and VocabFeatureExtractor. Linguistic feature-based models are compared head-to-head against transformer-based models (BETO, MarIA, BERTIN).

I also included six feature-selection methods voting on what actually matters, with collinearity pruning deliberately moved to after the tournament rather than before it. That said, the method wasn't more effective than L2 Synthesis feature selection.

~0.96
QWK · BERTIN + CatBoost hybrid
~87%
classification accuracy
822
texts, A1–C2
238
linguistic features

The best model pairs BERTIN with CatBoost — the "BERTIN paradox": weakest model alone, strongest as a hybrid partner. This model uses 69 features, selected with L2 Synthesis feature selection, not the full 238 from the initial pipeline. It also holds up well against MultiAzterTest, the existing benchmark, beating it on CEFR discrimination while correlating strongly with it (ρ = 0.738).

Study 2 — does difficulty feel the way the model says it should?

Study 2 triangulates the model's predictions against real perceptions from Mandarin-L1 learners of Spanish, including affective variables like anxiety and engagement, and typological distance. Including affect required making a literature-backed case strong enough to survive real methodological pushback — not just asserting it belonged.

Before the Ph.D. — the master's

My M.A. in Applied Foreign Languages (NTUST, 2020–2022), advised by Dra. Sy-Ying Lee, asked a narrower version of the same question: are graded readers or authentic illustrated books better first material for beginning Spanish learners? A corpus analysis of verb variety, frequency, mood, and tense across both text types, plus a 12-week study with Taiwanese Spanish majors, became a 170-page manuscript — and the seed of everything since.

Presented at Edge Hill University (UK), Teachers College Columbia University, Westminster International University in Tashkent, and National Taipei University of Business. Two more papers on Study 1 & 2 are queued for the XII Congreso Internacional de la AAH.

Alongside all this, a handful of smaller but important milestones pulled me further into the computational side: Text Analytics with Python (edX, 2023), a Professional Python Diploma from PUC (2023), CS50x: Introduction to Computer Science (2023), and CS50's Introduction to Programming with Python (HarvardX, 2026).

Publications & reviewing

Morphosyntactic Features for CEFR Text Classification of L2 Spanish

From my doctoral studies. In active review for the PACLIC Conference.

Spanish Graded Readers vs. Authentic Illustrated Texts

Co-authored with Sy-Ying Lee. In active revision for the journal System, after an earlier round at Reading in a Foreign Language.

Uso y adquisición de artículos en español como segunda lengua

Valenzuela & Toledo (2019). Logos: Revista de Lingüística, Filosofía y Literatura, 29(2), 268–285.

I also review for Revista Literatura y Lingüística (since 2022). From 2020 to 2026, until AI translation tools made the role redundant, I volunteered as a proofreader and English–Spanish translator for The Ehlers-Danlos Society, working on academic medical papers about EDS — see the honest part for what came after.

play

Go play

To game, or not to game: that is the question.

Time for a quick break — go play.

All games →

B2 projects

Things I've built

Research doesn't stay in papers if I can help it — most of it turns into a tool, a site, or an app. Grouped below the way I actually think about them: EdTech first, Health Tech after.

EdTech

The Carla Model

CARLA is a hybrid readability model for Spanish — the engine behind Study 1, built, trained, and shipped solo: fine-tuned open-source transformers (BERTIN, BETO, MarIA) on consumer hardware, fused with 238 psycholinguistic features (69 after L2 Synthesis selection) into a CatBoost ensemble, and released as a live web app you can paste any Spanish text into. It returns a CEFR level (A1–C2) at ~0.96 QWK / ~87% accuracy on 822 CEFR-labelled texts.

The "BERTIN paradox" is the frugal-engineering point: the weakest model alone became the strongest hybrid partner — no big lab, no huge GPU budget, just deliberate feature design. It's a working answer to the question a decade of teaching left me with. Is this text too hard for the person in front of me? Named for my 乾女兒 — my goddaughter — whose Spanish name is Carla, read as a backronym: Computational Assessment of Readability and Leveling Automation.

Where it's going: a recommender pairing CARLA's score with Project Gutenberg's Spanish catalog, so a learner at low B1 can find authentic texts they can actually read — not "graded" down to their level, not miles past it. Building CARLA also gave me the technical confidence to start ZebraUp. But this one came first. It's still the one I'm proudest of.

PythontransformersNLPBERTINCatBoostFeature EngineeringCEFRStreamlit

CKJ-ESP

CKJ-ESP is the idea that used to live a few sections down, listed as a hazy "shape TBD" — it shipped. A Spanish-first, Pleco-style comparative dictionary connecting Mandarin (Taiwan), Korean, and Japanese, side by side, in Spanish: vocabulary, grammar, and writing systems, plus the connections a "one language at a time" course never draws. Ten articles live so far — the copula (是 / 이다 / です), keyboard layouts (注音, hangul, かな), how to actually read hangul and hiragana/katakana, and the 214 Kangxi radicals among them — sourced from TOCFL/MOEdict, KRDICT, and JMdict, with stroke order from Taiwan's Ministry of Education. Built as a static site — no server, no tracking, just text and structure.

Where it's going: more articles, closer integration with Tonestamp and leehzn, and eventually an API for anyone building CJK ↔ Spanish tools. This is the language project I move between three languages for daily — which makes it the one I'm most personally invested in keeping alive.

Comparative LinguisticsLexicographySpanish ↔ CJKTOCFL / KRDICT / JMdictStatic Site

NAS Project

The answer to the name game. A system built to acknowledge and respect diverse name structures and preferences — because most software still assumes everyone's name fits one Western template.

github.com/Poln4/nas-project →
PythonInternationalization (i18n)Data ModelingInclusive Design

Reading Club

A reading-club dream going back to at least 2020, when I started collecting graded readers and reading them aloud with students, watching what it did for them firsthand. This site is where that idea actually lives.

Reading Club → @clubdelecturaele on Bluesky → Play Myth or Fact? →
FlaskPythonEdTechExtensive ReadingCommunity Building

Tone Stamp

A reading tool for learners of Chinese that colors characters to make components, radicals, or pronunciation patterns visible at a glance — the same instinct behind my Spanish readability work, aimed at a language I'm still learning myself. Now paired with Decks: free, browser-based flashcard sets organized by TOCFL level (準1 through 流利/Fluent), with flip cards, keyboard navigation, a zhuyin/pinyin/off reading toggle, and room for your own example sentences. Two of seven levels are live so far (準1 · 152 cards, 準2 · 179 cards); studying online stays free, with an optional Ko-fi link for anyone who wants to export or print a deck.

tonestamp.netlify.app → Decks →
EdTechChinese PedagogyReadabilityVisualizationFlashcardsTOCFL

Health Tech

ZebraUp

ZebraUp is an mHealth app for people navigating rare or hard-to-diagnose conditions — the "zebras," in medical slang, that get overlooked while clinicians look for more common explanations. It helps patients walk into short consultations with concrete data instead of a blurred memory of symptoms: structured symptom tracking, a timeline they can hand to a new specialist, and language built around how patients actually describe what's happening to them, not just clinical checklists.

It is also a study in solo, privacy-first engineering: a local-first Dart/Flutter app designed so the patient's data stays on their device, with no mandatory cloud dependency — a deliberate single-developer architecture that removes servers, subscriptions, and vulnerability surface from the equation. The point is legibility: making an invisible, disbelieved illness visible to the only person who can treat it.

I'm currently gathering structured data on how patients with complex, multi-system conditions communicate with their care teams — where that communication breaks down, and what actually helps — with the aim of publishing a study that uses ZebraUp as its case study. More on why I built this in the honest part.

DartFlutterhealth techmHealthPatient AdvocacyUX Writing

ZebraPace

🦓 "The 1% Journey." A pacing-first energy and fitness tracker for people managing Ehlers-Danlos Syndromes, Post-Exertional Malaise, dysautonomia, and their comorbidities. Traditional fitness apps reward streaks and linear progress — which can actively harm chronic, cyclical conditions — so ZebraPace works the other way: a "spoon" battery that reflects real daily capacity, a gentle 1% growth rule instead of a push, and PEM-precaution alerts if you overdo it. Rest counts as a successful health intervention, not a broken streak.

Grew directly out of the same story behind ZebraUp. Currently running with just me as the case study (early beta on Streamlit). If you're navigating something similar and are curious, I'm happy to share this first phase before building it out further.

zebrapace.streamlit.app →
early beta Health TechPacingEDS / PEMStreamlit

Not everything below has shipped yet. These are the ones still living mostly in my head.

? dream projects

Draft v.1.0 — in my head, often

Split the same way as Projects above. EdTech: leehzn, a graded Spanish-reading platform for Korean/Chinese/Japanese speakers, and El Vecindario, a visual-novel Spanish-learning game built around active production and zero-punishment failure. Health Tech: ZebraPace, a pacing-first energy and fitness tracker for EDS, PEM, and dysautonomia.

Ideas I keep coming back to, in various stages of "not real yet." Listed here so I stay honest about the gap between dreaming them and shipping them.

EdTech

leehzn

A graded Spanish-reading platform built specifically for Korean, Chinese, and Japanese speakers — contrastive grammar explained from each native language, levels calibrated with my own CARLA readability model, real Latin American Spanish from day one.

Leehzn →
not public yet

El Vecindario

A visual-novel Spanish-learning game designed against the "streak trap" — no health bars, no game over. Vocabulary is acquired by dragging mystery words onto context, not translation; sentences are built from word-blocks to force active production instead of multiple choice. Starts in a soft-landing family home (A1) and expands into a full neighborhood — market, deliveries, a lost-internet crisis — through A2.

idea stageGame DesignFailing ForwardSLA Pedagogy

All of the above runs on a fairly specific toolkit.

toolkit

What I build with

PythonRC / C++SQLHTML / CSS spaCyscikit-learnNLTKXGBoost transformers (BETO, MarIA, BERTIN)pandasNumPy corpus linguisticsFlaskGit / GitHub MoodleH5P

Soft skills

Detail-Oriented Multidisciplinary Self-Directed Cross-Cultural Communicator Resilient

Languages

Spanish — nativeEnglish — fluent (TOEFL iBT 102/120)Mandarin — intermediateKorean — elementary

Affiliations

LSA — Linguistic Society of America ASELE: enseñanza del español como lengua extranjera

None of this happened as smoothly as it reads so far. Here's the part usually left out.

C1 the honest part

The messy middle

6 scholarships took me from Chile to South Korea, New York, and Taiwan. Stopped competitive coaching due to Ehlers-Danlos Syndrome, and later saw AI make my Ehlers-Danlos Society volunteer translation work redundant — both experiences led directly to ZebraUp. Also navigating a job market where formal training competes with untrained native speakers and AI-assisted work.

A live CV usually hides the parts that weren't smooth. Here's some of what actually got me here.

Nobody in my family had a map for any of this, so I've drawn most of it myself, one scholarship application at a time. Six of them, so far, got me from Chile to South Korea to New York to Taiwan.

Chronic illness has its own tax here. Real medical support for a connective-tissue condition is hard to find anywhere, and harder still in Taiwan — finding specialists who actually understand it has been its own separate fight. Steady work hasn't been simple either: in Taiwan, any native Spanish speaker can end up teaching Spanish with no training or credentials required, and now, anyone with access to AI can write enough code to pass for the work being done. It's hard not to feel years of training counting for less every year.

For a while, coaching was part of the plan too. Then Ehlers-Danlos Syndrome made that impossible — a connective-tissue condition that, after years of frustration and one knee injury too many, meant I had to stop coaching soccer and basketball for good. I even trained and earned a Football Statistician Certificate (Genius Sport, 2024), to stay near the game from the sidelines and report matches around Taiwan — but my body wouldn't cooperate with that either. That long, exhausting fight to be heard about my own health is exactly what became ZebraUp. More recently, AI tools made my volunteer translation work for the Ehlers-Danlos Society redundant — that one stung, honestly — but it's part of what pushed me to put that energy fully into ZebraUp instead.

I also completed the Be There Certificate (Born This Way Foundation, 2023), on how to support someone through a mental health crisis.

And underneath all of it: I've been lucky to stretch a single five-year scholarship across two degrees, two years for the master's, three for the Ph.D., and I'm close to the finish line now.

C2 what's next

Work with me, or back this work

The invisible has texture — and I build lenses for it. Tone Stamp colors the tones of Mandarin you can't hear; ZebraUp draws the pain patterns of a chronic condition a doctor can't see; CARLA turns the hidden difficulty of a Spanish text into a number anyone can read. Same problem, same engineer. EdTech and health tech aren't two jobs — they're two applications of the same instinct: making the invisible legible.

An engineer of lenses: whether I'm coloring invisible Mandarin tones or charting invisible illness symptoms, the work is the same — turning complex, unseen reality into something clear and human.

Why back this work

Institutions with millions in funding haven't built tools like these for language learners or chronic-illness patients. A solo developer did, on a shoestring, with zero operating cost. Funding this work isn't charity — it's funding R&D for communities the system overlooks, run by the person who already knows how to ship. A small monthly contribution is what keeps the lenses free for everyone who needs them.

Get in touch Support on Ko-fi GitHub Sponsors Download CV ↓