Open Source — Proof by Construction
23 public repositories. One discipline: I build machine learning from scratch so I can see inside every model — and then I share the code. The public arc runs from-scratch → federated → applied & scientific.
One account, one through-line.
Everything on github.com/avilaqba is built from the inside out. I don't just call sklearn and move on — I write the math, the gradient, the optimization loop, the decision boundary, by hand, so that when I later build on top of these primitives (federated training, applied forecasting, connectomics), I trust what's under the hood. The 23 repos below are the public trail of that habit, in three phases.
From Scratch
The foundations. Regularized linear models, optimization loops, and a probabilistic classifier — all written from first principles, no off-the-shelf estimation.
Ridge · Lasso · ElasticNet — From Scratch
Regularized linear models written line-by-line — the core math implemented by hand.
Gradient Descent vs SGD — From Scratch
The optimizer battle for linear regression, implemented by hand to show the bias/variance computation trade-off.
Binary Bayes Classifier — From Scratch
A probabilistic classifier built from the fundamentals of Bayesian inference.
Distributed & Applied
From single models to coordination across data silos — then the applied forecasting and predictive modeling that mirrors my commercial data-science work.
Federated Learning with Logistic Regression
Private, distributed training across data silos without moving the data — the on-ramp to distributed learning.
Bank Customer Churn Prediction
Applied predictive modeling on a real business problem — the notebook-to-service mindset.
Forecasting — Bus Demand · Solar Power · Incoming Calls
Time-series prediction across three domains: bus demand, solar-power generation, and incoming-call volume.
Scientific
Where the thread reaches biology and emergent computation — connectomics, molecular-graph theory, and population-based search.
Brain Connectivity Matrix Super-Resolution
Recovering high-resolution brain wiring graphs from coarse connectomes — released as a CC0 Kaggle dataset (289 downloads).
Morgan Algorithm — Unique Atom Numbering
Graph-theoretic canonical numbering of atoms in molecules — chemistry + graph theory + code.
TSP with a Genetic Algorithm
Combinatorial optimization via evolutionary search — population-based, emergent problem solving before I did swarms.
An honest note on the LLM & agentic work.
My recent work — LLM applications and agentic AI systems, orchestration, tool use, memory, and evaluation — is professional work done at Property Turkey, and it is not yet open-sourced on this account. This page is a deliberate record of what is public: a trajectory that starts with building ML from scratch, moves through federated/distributed learning, and lands in applied and scientific work. The public code shows the foundation the agent work stands on — it doesn't overclaim the agent work itself as open source.