What I can do,
and what I'm learning.

Python

Feature engineering, Random Forests, CORAL adaptation, and evaluation scripts for Cross-City Building Age Classification; AfriGuard and Digital Minds pipelines.

TypeScript / JavaScript

This website and SA Data Hub, both typed React/Next.js applications in production.

Supervised classification

Four-class building-age prediction from Landsat features; stylometric author identification baselines in Digital Minds.

Random Forests & scikit-learn

Class-balanced RF (500 trees) as Stage-1 Madrid model; local few-shot forests in the Hack4Dev transfer pipeline.

Feature engineering

~108 per-pixel features from multi-period spectral statistics, indices, change timing, and neighbourhood averages.

Domain adaptation (CORAL)

Pooled and class-conditional CORAL for Madrid→Amsterdam zero-shot transfer (0.36 → 0.65 macro-F1).

Model evaluation & cross-validation

Madrid 5×5 CV, macro-F1 transfer curves, McNemar tests, and capability-controlled 2×2 designs across 9,269 LLM trials.

Spatial analysis

Neighbourhood features, inverse-distance smoothing, and a spatial-adjacency audit of few-shot support/query leakage.

ETL & automated processing

AfriGuard end-to-end evaluation pipeline (1,120 responses); SA Data Hub automated update scripts; Landsat parquet loading and feature builds.

Pandas & NumPy

Core stack for AfriGuard analytics, Digital Minds scoring logs, and the building-age feature matrix.

Public-data ingestion

SA Data Hub ingests and documents Stats SA, SARB, and other official South African sources with methodology pages.

Data validation & reproducibility

Fixed seeds, logged API calls with prompt hashes, and deliverable tables/figures regenerated from notebooks.

Multilingual AI safety evaluation

AfriGuard: 7 languages, 4 frontier models, ASR analysis showing guardrail collapse outside English.

LLM introspection research

Digital Minds: capability-controlled self-prediction probes versus cheap surface baselines.

Experiment design & baselines

Crossed 2×2 designs, length-only and stylometric comparators, and organiser-notebook baselines for transfer.

Bias & leakage diagnostics

Surface-leakage gate, response-bias checks, and spatial-adjacency audits of few-shot evaluation.

Next.js & React

Personal site (ubayd.me) and SA Data Hub (sadatahub.tech).

Tailwind CSS

Responsive UI systems across both sites, including dark/light themes.

Recharts

Interactive visualisations on SA Data Hub dashboards and explorers.

Streamlit

AfriGuard live analytics dashboard for language and harm-category exploration.

Vercel

Hosting and deployment for this website and SA Data Hub.

Git & GitHub

Version control across AfriGuard, Hack4Dev, Digital Minds, and personal projects.

Command line

Running transfer notebooks, evaluation scripts, and local build/deploy workflows.

VS Code

Primary development environment for Python and TypeScript work.

APIs

OpenRouter and frontier-model APIs for Digital Minds and AfriGuard evaluation collection.

SQL & PostgreSQL

Learning relational querying and database design for a future SA Data Hub backend.

Algorithms & data structures

Strengthening foundations through UCT Computer Science coursework.