Open to internships · 2027

Süleyman Kılınç

Software engineering student

Location:Elazığ, TürkiyeCzęstochowa, PolandErasmus+ from February 2027

Second-year software engineering student at Fırat University. I build software and measure it: a frame-rate model checked against real benchmarks, and a voice assistant that runs entirely on-device.

Fırat University · 2nd year
Erasmus+ · Politechnika Częstochowska
English · B2

View projects Email me LinkedIn GitHub

Selected work

Featured projects

For each one: the problem, what I built, and what I measured.

Web app · Performance model

PerfHub

PC games covered: 176; PC games covered
recorded benchmarks: 741; recorded benchmarks
mean error on fitted runs: 6.7%; mean error on fitted runs

perfhub.suleymankilinc.com

PerfHub results page for a Ryzen 5 7600 with an RTX 4070 at 1440p

Problem: Before buying a part, people want to know what frame rate they'll get. Published benchmarks cover only a few CPU and GPU pairs, and online calculators rarely say how wrong they can be.
What I built: A frame-time model in which the slower of the CPU and GPU sets the pace, plus a separate VRAM model, fitted to recorded benchmarks and running entirely in the browser. A desktop app in the same repository detects your hardware automatically and is being moved onto the same engine.
What I measured: 6.7% mean error on 741 recorded benchmark runs, with 78% of predictions within 10%. Games that were never measured are tested by holding them out (about 52% off) and are labelled as estimates. Every figure is regenerated from the data, and CI fails if one goes stale.

Python
TypeScript
GitHub Actions

Try the live demo Source code

Desktop · Local AI

Halcyonformerly Jarvis

voice-to-voice response on GPU: 1.1–1.6 s; voice-to-voice response on GPU
on-device speech pipeline: 100%; on-device speech pipeline
language detected per utterance: TR · EN; language detected per utterance

Pipeline, measured on an RTX 4060 Laptop GPU

1
Microphone
16 kHz, 32 ms frames
2
Voice detection
Silero VAD · 480 ms of silence ends a turn
3
Speech to text
faster-whisper · 234–377 ms
4
Local language model
qwen3.5 via Ollama · first token in 114–134 ms
5
Speech
Piper · starts with the first finished sentence

First audio after 1.1–1.6 s

Problem: Voice assistants usually send your audio to the cloud, and local ones tend to feel slow, because you wait for the whole answer before hearing anything.
What I built: A streaming pipeline: voice activity detection, faster-whisper, a local qwen3.5 model through Ollama, and Piper speech synthesis. It speaks the first sentence while the rest is still being generated, can be interrupted mid-sentence, and picks models that fit the GPU's free memory, then checks that nothing spilled onto the CPU.
What I measured: 1.1–1.6 s from the end of your sentence to the first sound back, on an RTX 4060 Laptop GPU. It's measured with a benchmark that feeds synthesised speech through the real pipeline; making that benchmark reproducible meant pinning three hidden sources of randomness.

Python
Ollama
faster-whisper
Piper TTS
PyTorch · CUDA

Source code

About

Where I am, and where I'm going

I like problems where the answer can be checked. When I build something, I want to know how well it actually works, so I measure it, write down where it falls short, and fix the part that's really broken. I'm looking for a team where that habit is useful and where I can learn from engineers who've shipped real systems.

What I'm looking for

A software engineering internship in backend, developer tooling or applied AI. In the long run, I want to build my career in Europe.

Availability

Part-time: from February 2027, alongside my Erasmus+ semester
Full-time: summer 2027
Hybrid: Częstochowa and nearby cities
Remote: teams anywhere else

2025 – present
BSc Software Engineering
Fırat University, Elazığ, Türkiye
Feb 2027 – Jun 2027
Erasmus+ exchange semester
Politechnika Częstochowska, Poland

Skills

Tools I work with

Languages

Python
Java
C#
TypeScript

Web

React
Next.js
Flask

Data and AI

PostgreSQL
Local LLMs (Ollama)
Transfer learning

Tooling

Docker
Git

Spoken languagesTurkish · nativeEnglish · B2

How I use AI

Most of the code in my recent projects was written together with Claude Code, an AI coding assistant. I'd rather tell you up front than have you find it in the commit history.

My part

I choose the problem, decide what "working" means, review every change and decide what ships. I run the measurements myself, and a number goes into a README only after I've reproduced it.

Its part

Drafting code, tests and documentation, and getting me productive in unfamiliar libraries much faster than I could manage alone.

Where I don't trust it

Plausible output isn't evidence. In Halcyon, an early benchmark returned neatly consistent numbers that turned out to be measuring its own backlog. Catching that kind of thing is the part I don't hand off.

What I build without it

Fundamentals: algorithms, data structures, OOP and SQL. I practise them by hand in coursework and in the small Java projects on my GitHub.

Tools

Claude Code · AI coding assistant in my projects
Ollama · running language models locally in Halcyon

You can check this in the commit history: commits written with the assistant carry a "Co-Authored-By" line. Commit history: Halcyon, PerfHub

Blog

All posts →

Contact

Let's talk

Hiring an intern for 2027, or want to talk about one of my projects? Email is the fastest way to reach me.

kilincsuleyman.dev@gmail.com

Send an email LinkedIn GitHub