Selected work, organized by what it’s trying to do.

11 case studies 8 live products 41 technologies Start with research → See the products

★ Featured project

Glioma Subtype Classification & Survival Modeling from DNA Methylation

Classifying brain-tumor subtypes and predicting survival from DNA methylation.

97–99%Subtype classification accuracy

0.796Survival model C-index

404K → 70Features compressed

PythonRXGBoostglmnet (Elastic-Net Cox)CoxBoostscikit-learn

Read case study →

Health & ML research 4 projects

Machine learning and bioinformatics applied to clinical and biological problems.

Predicting depression severity from routine EHR data.

ProblemMeasuring depression severity normally needs a separate PHQ-9 visit, so patients often get flagged late.

My roleLed a UBC MDS capstone for GreenspaceHealth: engineered features from 140K+ multi-clinic encounters and built a calibrated, explainable model that predicts PHQ-9 severity without the questionnaire.

OutcomeShipped a deployable model with feature-importance reports — showing routine EHR data alone can flag depression between visits.

140K+ EHR encounters processed

MDS Capstone project, UBC

Multi-clinic Generalized across sites

Python, Snowflake SQL, scikit-learn, pandas, Random Forest, Temporal Neural Network

Read case study →

Predicting PHQ-9 Depression Scores from Multi-Clinic EHR Data

Benchmarking transfer-learning architectures on chest X-ray classification

ProblemHospital radiology departments increasingly want to test off-the-shelf deep models on chest X-rays, but accuracy comparisons across architectures are scattered across papers with different datasets and metrics.

My roleBenchmarked four pretrained CNN architectures (ResNet-50, DenseNet-121, EfficientNet-B0, VGG-16) on a unified chest X-ray dataset, controlling for input pipeline, fine-tuning regimen, and evaluation protocol. Measured accuracy, AUC, sensitivity, specificity, and inference time per architecture.

OutcomeApples-to-apples comparison showing which architectures actually generalize on lung-disease X-rays under matched conditions. Reproducible notebook with full benchmarking pipeline.

4 archs Apples-to-apples benchmark

Unified Single dataset + metrics

Reproducible Open notebook

Python, TensorFlow, Keras, Matplotlib

Read case study → Live demo Code

Comparative Evaluation of Pretrained Models for Lung Disease Classification

Predicting diabetes risk from clinical features with multiple ML approaches

ProblemPublic diabetes-prediction tutorials usually report a single accuracy number on a single split — useless for understanding when a model will fail in deployment.

My roleBuilt a comparative diabetes risk pipeline (logistic regression, random forest, gradient boosting) with stratified cross-validation, threshold tuning, and per-class performance breakdown. Surfaced where each model under- and over-predicts across the population.

OutcomeCalibrated risk model with explainable feature importance and confusion-matrix breakdown by demographic. Shows the trade-offs between sensitivity and specificity that matter for actual clinical screening.

3 models Compared head-to-head

Stratified Cross-validated

Calibrated Threshold-tuned per use case

Python, Scikit-learn, Pandas, Matplotlib, Seaborn

Read case study → Live demo Code

Classifying milk grade from physicochemical measurements

ProblemDairy QC labs catch contamination at the end of the line, after bad batches are already produced. Earlier prediction from upstream sensor data would catch issues before a batch is ruined.

My roleBuilt a multi-class classifier predicting milk quality grade (low / medium / high) from chemical sensor features (pH, fat, protein, lactose, somatic cell count). Compared tree ensembles against linear baselines and tuned class weights to handle imbalanced grade distributions.

OutcomePredictive model identifying milk grade from upstream sensor data alone, with feature importance pinpointing which chemical signals matter most. Useful pattern for any process-quality prediction problem.

Multi-class Low / med / high grade

Sensor-only No downstream lab tests needed

Imbalanced Class-weighted training

Python, Scikit-learn, Pandas, Matplotlib, Seaborn

Read case study → Live demo Code

Products I ship 3 projects

Indie products designed, built, deployed, and operated end-to-end.

Daily 60-second coding and data-science lessons, built solo.

ProblemMost coding courses are too long, so learners quit before they finish anything.

My roleDesigned, built, and operate the whole platform solo — a multi-tenant Django app serving daily micro-lessons with Stripe subscriptions, progress tracking, and exercise grading. One brand-aware backend powers both BitSize Learning and BitSize Language.

OutcomeLive at bitsizelearning.com with an active paid tier. Every line of code, deploy, and migration is mine.

Solo Founder, engineer, operator

2 brands One backend, two products

Live Stripe-paid subscribers

Django, PostgreSQL, Cloudflare R2, Stripe, SendGrid, Google OAuth, Railway

Read case study → Live demo

Daily Japanese & French lessons with speech-recognition practice.

ProblemLanguage apps reward streaks, not speaking — so learners can read but freeze when they have to talk.

My roleSecond brand on the BitSize platform: daily Japanese and French lessons with in-browser speech recognition for pronunciation, native TTS audio, and a JLPT N5–N1 grammar module.

OutcomeLive at bitsizelanguage.com — real-time pronunciation feedback in the browser, no app install.

2 langs Japanese + French

N5–N1 Full JLPT coverage

Browser Speech recognition, no app

Django, PostgreSQL, Web Speech API, Edge TTS, Cloudflare R2, Stripe, Railway

Read case study → Live demo

This portfolio site — designed, built, and deployed solo on Django + Render

ProblemMost data scientists either link a Squarespace site or push to a free GitHub Pages repo. Neither communicates 'I can ship and operate production systems' — which is exactly what hiring managers screen for.

My roleDesigned and built this portfolio as its own production system: 7 Django apps, PostgreSQL on Render, WhiteNoise compressed manifests, Cloudflare R2 for media, allauth Google OAuth, Stripe ready, deploy-on-push CI/CD. Embedded the Spotify Explorer dashboard directly as a Django app (no external Render service) and built a CMS-grade admin so all content edits flow through Django migrations or admin, not template hacks.

OutcomeLive at nijiati-abulizi.com — every page, every chart, every API endpoint runs in this single Django process. The portfolio is itself the strongest case study in the portfolio.

7 apps Django modules

1 process All charts self-hosted

Solo Designed, built, operated

Django, Python, HTML, CSS, JavaScript, PostgreSQL, Bootstrap

Read case study → Live demo Code

Tools & open source 3 projects

Dashboards, packages, and data tools for analysts and researchers.

An interactive dashboard for exploring music taste and trends.

ProblemSpotify has rich audio data, but no easy way to explore your own taste or find similar tracks.

My roleBuilt a three-tab Plotly dashboard — explore your taste, find new music, and map global trends — in pure pandas + Plotly.js, embedded right in this site.

OutcomeSelf-hosted at /projects/spotify-explorer/ — loads instantly, with no external service or cold starts.

3 tabs Taste / Discover / World

Self-hosted No external service

Instant No Render cold start

Python, Plotly Dash, pandas, Spotify Web API, Render

Read case study → Live demo Code

An R package for Yelp business analytics and maps.

ProblemYelp hides location patterns in long lists — a map shows where a business works at a glance.

My roleBuilt a reusable R package wrapping the Yelp API: choropleth maps, density heatmaps, and rating comparisons, with CI and 88% test coverage.

OutcomeA published, reusable R package for spatial business analysis — useful for retail, real estate, or delivery.

R package Reusable across datasets

Spatial Choropleth + heatmap

2 install paths (devtools::install_github + .tar.gz)

R, Yelp Fusion API, Plotly, ggplot2, devtools, GitHub Actions, testthat

Read case study → Code

Pip-installable Python quiz app, packaged and distributed via PyPI workflow

ProblemStudying for technical exams means re-reading the same material until it sticks. Active recall works better, but most quiz tools are tied to a specific course or platform.

My roleBuilt a self-hosted quiz application supporting custom question banks, randomized order, and immediate feedback. Backend in Django, mobile-friendly UI, exportable progress reports.

OutcomePersonal study tool used to drill exam material with active recall. Pattern reusable for any quiz-based learning workflow.

1 pip-installable command-line entry point

src layout with pyproject.toml + setuptools

End-to-end from source code to installable distribution

Python

Read case study → Live demo Code