Tekin Ertekin tekinertekin

Tekin Ertekin

Software Engineer — 12+ years building production systems. Currently working on low-precision numerics and machine-learning infrastructure in C.

Open-source contributions

15+ merged pull requests across four organisations. Click a logo to open that organisation's list of my merged pull requests.

NASA
_{F′ & cFS flight software · 4 merged}

ROOT
_{CERN · 5 merged}

SciPy
_{scipy & xsf · 4 merged}

llama.cpp
_{ggml-org · 2 merged}

Mostly numerical correctness and performance work: premature overflow in Bessel functions, a flood fill that hung or leaked past its outline depending on the brush alpha, a socket timeout that was configured but never applied, generation settings an inference server accepted but never reported back, and a transposed convolution that silently computed only the first batch.

mantissa

mantissa — a low-precision numerics core for neural networks, written in C.

It stores weights and activations in narrow formats (fp8, fp4, and a block-scaled variant) and runs dense and convolutional layers over them with hand-written NEON kernels. Every performance claim in the repository is backed by a paired randomised measurement against a same-binary control, and results that did not survive that test are documented as rejected rather than removed.

rubicon

rubicon — a small agent that decides its own steps and escalates to a human when it isn't sure, written in Python on LangGraph.

It classifies an arXiv paper into its subject area from the abstract alone, over real arXiv data: it retrieves similar papers, judges its own confidence, and then chooses what to do next — commit the answer, gather more evidence, or hand the case to a person. On the confident band it decides on its own at 88.1% accuracy and routes the uncertain ~30% to a human. Wiring a local 7B model to override the confident picks lowered accuracy, so that path is kept opt-in and the result is documented rather than dropped — the same paired-measurement discipline as mantissa.

Education

	PhD, Computer Engineering — Hacettepe University, 2021 – present _{Thesis in progress. First paper published in Applied Sciences, 2026 — Concept Tree Learner (CTL).}
	MSc, Game Technologies and Computer Animation — Hacettepe University, 2021 _{adaptive-voronoi-mapping: adaptive Voronoi mapping for legible visualisation of very large labelled spatial point data.}
	BSc, Computer Engineering — Middle East Technical University, 2014

Publications

M. T. Ertekin and B. Genç, "Concept Tree Learner (CTL): An Incremental and Interpretable Symbolic Framework for Binary String Rule Induction", Applied Sciences, vol. 16, no. 17, art. 8542, 2026. Open access, doi:10.3390/app16178542
M. T. Ertekin and B. Genç, "User Oriented Visualization of Very Large Spatial Data with Adaptive Voronoi Mapping (AVM)", in Engineering Cyber-Physical Systems and Critical Infrastructures, Springer, 2023, pp. 515–536. doi:10.1007/978-3-031-31956-3_45
M. T. Ertekin, "Çok Büyük Konumsal Verinin Kullanıcı Odaklı Görselleştirilmesi", MSc thesis, Institute of Informatics, Hacettepe University, 2020. In Turkish, full text

Work

Software Engineer.

Before that, national-scale public-sector systems in .NET — both still in production:

KYK, Turkey's student loans and housing agency — worked on a team that built 9 databases and 11 applications from scratch. Around 15 million users.
Turkish Employment Agency (İŞKUR) — employment and public-benefit programme systems serving 8+ million users.

Other work

Flight-controller firmware — modified the C sources running on a Pixhawk Cube autopilot.

Navigation Menu