// LLM & AGENTIC AI ENGINEER · ISTANBUL, TR

Behzat Aslanoğlu

80citations 4h-index 23open-source repos 10certifications 289dataset downloads 3.94MSc GPA

The story, the systems, the science — in one place.

01 · mathematician 02 · scientist 03 · builder 04 · researcher 05 · now & next 06 · collaborate

01 — the mathematician

// foundationsIt starts with mathematics and code — together.

Mathematics, code, computer science, and the web grew up in my head together — each sharpening the others. I earned a BSc in Applied Mathematics (University of Tabriz), an MSc in Applied Informatics at Istanbul Technical University (3.94/4.00 GPA), and a PhD candidacy in Information & Communication Engineering there — while simultaneously teaching myself to program and to think in machine learning. The math informed the code; the code made the math real.

The discipline was never about memorising formulas — it was about decomposing a messy, real problem into equations you can actually solve. Differential equations, linear algebra, and probability became the grammar I read every dataset and network in, while the compiler was the pencil I used to test the sentences. That single thread — from abstract proof to working software — runs from molecular simulations to LLM agents.

The web entered alongside all of it. Designing for the browser taught me structure, systems, and how to communicate complex ideas — the interface layer of every data product and agent I later shipped.

None of it waited its turn. I reached for Python and the browser while the calculus was still warm, and taught myself machine learning the same season I first touched a debugger. The same years that produced an MSc and PhD candidacy also produced ten IBM and Kaggle certifications — earned as part of the study itself, never a tidy “later.”

education

PhD (candidate) — Information & Communication Eng., ITU
MSc — Applied Informatics, ITU · GPA 3.94/4.00
BSc — Applied Mathematics, Univ. of Tabriz

milestones

BSc Applied Math — University of Tabriz
MSc Applied Informatics — ITU · GPA 3.94
PhD candidate — Information & Communication Eng.
First code — the algorithm that made the math real
First website — the web as a creative canvas
First model — machine learning, learned in parallel

formerly known as

Behzad Aslani Avilaq — published research & open-source under this name.

the through-line

Math never waited for the code, and the code never waited for the math. They arrived together and stayed together — one continuous, intertwined thread from the very start.

the path — degree by degree

BSc

Applied Mathematics — University of Tabriz. Proofs, systems, and a lifelong love of structure.

MSc

Applied Informatics — Istanbul Technical University · GPA 3.94/4.00. The moment the math learned to compute.

PhD

Information & Communication Engineering — ITU, candidate. Deep research at the frontier.

Now

Coding, the web, and AI grew in parallel with every one of these — never after.

02 — the scientist

// researchInto biology: molecules & disease.

At TÜBİTAK I crossed from applied math into bioinformatics and molecular biology. My focus was the F876L point mutation — a single residue substitution that turns the androgen receptor into a resistance mechanism: it breaks enzalutamide's activity, converting the drug from inhibitor to agonist and driving prostate-cancer relapse. To understand it, I ran one-microsecond molecular dynamics simulations — stepping the forces on tens of thousands of atoms through time — which became a meeting abstract in the Biophysical Journal.

That work changed how I see data forever. Biology taught me to treat data as a physical system: every observation carries forces, constraints, and time-dynamics underneath. Just as a mutation only makes sense inside a folded protein, a number only makes sense inside the system that produced it. Respect the physics, model the mechanism — don't just fit the curve.

Because the enzalutamide–receptor complex was never experimentally solved, I docked the drug first, then ran one-microsecond simulations across the wild-type receptor, the testosterone-bound form, and several resistance mutants. The culprit: an unbending of helix 11 that strips stability from helix 12 — the switch that flips the drug from inhibitor to agonist. Add F876L, and that fatal unbending becomes alarmingly frequent.

research highlight

F876L mutation — a single residue substitution that flips enzalutamide from inhibitor to agonist, driving drug-resistant prostate-cancer relapse.

method

One-microsecond MD simulations — tens of thousands of atoms stepped through time, watching resistance emerge residue by residue.

venue

Biophysical Journal — investigation of enzalutamide drug-resistance mechanisms in the antiandrogen treatment of prostate cancer.

the path — from lab to frontier

2016

TÜBİTAK — bioinformatics & molecular dynamics; the F876L resistance story.

2018

Vodafone Future Lab — data science on telecom-scale, real-world data.

2021

Majorel Türkiye — senior data scientist on large customer-experience datasets.

2024

Property Turkey — LLM & agentic AI engineer, building agents and swarms.

03 — the neuroscientist

// brainThen the brain itself.

From molecules to neurons: my research turned to the most complex known graph — the human brain. I work on brain connectivity, building models that predict high-resolution brain graphs from low-resolution ones — super-resolution for connectomics. Where most imaging is too coarse to see fine-grained wiring, the goal is to recover the structure that is actually there: inferring dense, precise neural connectivity from sparse, noisy observations.

And, in the spirit of the science itself, I shared it openly. The brain connectivity matrices are published as a public Kaggle dataset so the neuroscience community can build on them — a small contribution to the idea that data, like brains, grows more powerful the more it is connected.

Super-resolution is not guesswork — it is constraint recovery. Given a coarse connectome, the models learn the anatomy and physics that govern real neural wiring, then fill in the missing density the way the brain itself would. It is the same instinct I carried out of molecular simulation: don't just fit the data — recover the system that generated it.

kaggle dataset

Brain Connectivity Matrix Dataset

Predict high-resolution brain graphs from low-resolution ones — 289 downloads, CC0, neuroscience.

kaggle.com/avilaqba →

milestones

289 downloads — and counting, from researchers worldwide
CC0 license — released into the public domain
Open science — the community keeps building on it
Super-resolution — predicting high-res brain graphs from low-res ones

04 — the machine learning era

// intelligenceTeaching machines to think.

Machine learning, built from first principles. Before the libraries made it easy, I built the algorithms by hand — Ridge, Lasso, and ElasticNet regularisation, logistic regression, the mathematics of gradient descent versus stochastic gradient descent, and the probabilistic foundations of a Bayes classifier. Writing an optimiser line by line is the best way to learn why it works; it also gave me the confidence to trust a model only when I could see inside it.

That from-scratch habit became a discipline of its own: a public GitHub library that grew from scratch implementations of machine learning into federated learning, then into agents — 23 repositories and counting. The credentials followed the craft: IBM certifications in Machine Learning with Python, Deep Learning Fundamentals, and Scalable Machine Learning on Apache Spark.

And then it all went into production: building and shipping ML systems across telecom, finance, and health at Vodafone Future Lab and Majorel — where the work stops being a notebook and becomes a service people rely on every day.

The thread is wider than any single field. My published work spans decision science and optimisation — balanced-scorecard prioritisation under uncertainty, fuzzy-TOPSIS risk assessment for EPC projects, warehouse-location ranking — and biomedical AI, from a computer-aided fuzzy expert system for kidney disease to the molecular-dynamics work. Different problems, one through-line: reduce a messy, real system to a model you can trust.

0citations

0h-index

0i10-index

from scratch · github

23 repositories — from-scratch ML → federated learning → agents. Built line by line to see inside every model.

certifications · IBM

Three IBM certifications — Machine Learning with Python, Deep Learning Fundamentals, and Scalable Machine Learning on Apache Spark.

certifications — a parallel transcript

Earned alongside the degrees, not after them — machine-learning, deep-learning, and data-engineering credentials from IBM, Kaggle, and Johns Hopkins, most issued in Dec 2020 – Jan 2021.

Machine Learning with Python — IBM · Dec 2020
Python for Data Science & AI — IBM · Dec 2020
Deep Learning Fundamentals — IBM · Jan 2021
Scalable ML on Apache Spark — IBM · Jan 2021
Fundamentals of Scalable Data Science — IBM · Jan 2021
Introduction to Data Analytics — IBM · Jan 2021
Tools for Data Science — IBM · Jan 2021
Big Data 101 — IBM · Jan 2021
Data Cleaning — Kaggle · Jan 2021
The Unix Workbench — Johns Hopkins Univ. · Jan 2021

05 — now & next

// frontierThe agentic era: LLMs & swarms.

Large language models, built in depth. Agentic AI that doesn't just answer — it acts. The frontier moved from models that predict the next token to agents that plan, call tools, and execute real work in the world. I build these systems end to end: orchestrating language models, tool use, memory, and feedback loops into something that can actually get a job done. Today that is my full-time craft as an LLM & Agentic AI Engineer.

And beyond single agents, AI swarms: many agents, coordinated. When independent agents share goals, decompose tasks, and negotiate solutions together, the whole becomes smarter than any one of its parts — emergent behaviour that echoes the very brains I studied earlier. Swarm intelligence means resilience and scale: no single point of failure, parallel exploration, and answers that come from collective reasoning rather than one model's guess.

An agent is a loop with a purpose. Around a language model I build the full scaffolding — the tools it can call, memory that persists across turns, planning that breaks a goal into steps, and a feedback loop that verifies the work actually happened. The result is software that doesn't just generate text; it performs tasks — the same instinct that earlier made me ship running systems instead of notebooks.

the craft · llm & agentic ai

Building agents — and the brains behind them

Today I build LLM systems end to end at Property Turkey: orchestration, tool use, memory, and feedback loops composed into agents that get real work done. Beyond single agents, my working thesis is the swarm — many agents, coordinated — where emergence echoes the brains I studied from molecules to connectomes.

orchestration · tools · memory swarm thesis — many agents, coordinated

connect with me →

It is the natural conclusion of a journey that began with mathematics and code side by side, moved through molecules and minds, and now builds intelligence that cooperates — the same intertwined thread, still unbroken.

Let's build the next layer of intelligence — together.

Mathematics, computer science, biology, neuroscience, machine learning — and now the frontier: LLMs, agentic AI, and swarms of agents. Every layer was learned in parallel, each reinforcing the others, and together they meet at one place: the work. If you have a problem worth solving, an idea worth building, or a team that needs an edge — there's a lot we can do together.

LinkedIn GitHub Google Scholar Kaggle avilaq16@itu.edu.tr