Proof by Construction.
The public record: from-scratch machine learning, federated learning, and a connectomics flagship that ties biology to the swarm thesis. I build line by line — and then I share the code.
The project that started the thesis.
Brain Connectivity Matrix Super-Resolution
Recovering high-resolution brain wiring graphs from coarse connectomes — super-resolution for neuroscience.
Builds models that predict dense, precise neural-connectivity matrices from sparse, low-resolution observations; released the matrices as a public CC0 Kaggle dataset (289 downloads) so the community can build on them.
Recovering fine-grained wiring from coarse scans is the methodological bridge from neuroscience to multi-agent intelligence — and the dataset is public so the field can build on it.
Beyond research: shipped products.
Two product lines engineered end to end under the Behzat AI banner — one for researchers, one that lives in your pocket.
Microx
The research workspace that sees the whole picture — consensus across AI models, expert reports and source verification, on one canvas.
open microxMneme
A private, on-device AI twin — it remembers your life, health signals, conversations and commitments in one searchable mind.
open mnemeThe from-scratch library & applied work.
Federated Learning with Logistic Regression
Private, distributed learning — training across data silos without moving the data.
A federated-learning implementation with a logistic-regression classifier: the on-ramp from classic ML to privacy-preserving, distributed training.
Training across data silos without centralizing the data — distributed coordination that answers real privacy constraints.
Ridge · Lasso · ElasticNet — From Scratch
Regularized linear models written line-by-line from first principles.
From-scratch implementations of Ridge, Lasso and ElasticNet regression — no libraries for the core math.
Built from scratch — Ridge, Lasso, ElasticNet, gradient descent, Bayes — so every model is inspectable.
Gradient Descent vs SGD — From Scratch
The optimizer battle, implemented by hand for linear regression.
Side-by-side from-scratch gradient-descent vs stochastic-gradient-descent implementations, showing the bias/variance computation trade-off.
Optimization is the micro-mechanism of learning — the update rule that decides how a model actually learns.
Binary Bayes Classifier — From Scratch
A probabilistic classifier built from the fundamentals of Bayesian inference.
Binary Bayes classifier written from scratch — no off-the-shelf estimation.
TSP with a Genetic Algorithm
Combinatorial optimization via evolutionary search.
Solves the Traveling Salesman Problem using a genetic algorithm — the earliest public hint of population-based, emergent problem solving.
Evolutionary and population-based search — my first love in AI, still my favorite tool for hard optimizations.
Bank Customer Churn Prediction
Applied predictive modeling on a real business problem.
Customer-churn prediction on banking data — representative of my commercial data-science work.
Predictive modeling as production work — the notebook-to-service path that mirrors real deployment.
Morgan Algorithm — Unique Atom Numbering
Graph-theoretic canonical numbering of atoms in molecules.
Implements the Morgan unique-atom-numbering algorithm for molecular graphs — chemistry + graph theory + code.
Graph-theoretic chemistry — canonical atom numbering that ties molecular structure to computation.
Every one of these was built from the inside out — I trust a model only when I can see inside it. That discipline is what I bring to agent systems: orchestration, tool use, memory, and evaluation, all built on the same implementation depth.