Portfolio · Selected work, 2026
AI Engineer. LLM systems, computer vision & data.
I build AI systems that run in production: retrieval and agent pipelines on large language models, computer vision for manufacturing, and the data work in between.
Three years at Bosch on vision inspection and process automation, then production RAG and GraphRAG systems at RADIAITE and Hagia. I study AI Engineering for Autonomous Systems at TH Ingolstadt and write framework-free courses on RAG, agents and evals.
Open source · Web · Desktop · Command line
Software I build and run, all of it open source.
Oct 2026 · MIT
Open-source live polls, word clouds and open questions for talks, classes and meetings.
Start a presentation and add your questions: multiple choice, word cloud, open question or scale.
Put the presenter screen up. People open your Oyla server on their phones and type the 6-digit code, or scan the QR code.
Votes, words and answers show up on the big screen live while people respond.
More products
Desktop app · Markdown · Tauri
A small notebook for Linux, Windows and macOS: press a shortcut, write the idea down, press Enter, and you're back to work.
Oct 2026 · MIT
Command line · Go · Coding agents
Connects EVREN, Türkiye's national model API, to the coding tools you already use: Qwen Code, opencode, Codex CLI and Claude Code.
Oct 2026 · AGPL-3.0-only
Claude Code hooks · Shell script · No dependencies
Plays a sound when Claude Code finishes a long turn, and stays quiet when you are already looking at the terminal.
Jul 2026 · MIT
3 courses · 57 essays · English & Turkish
Framework-free courses that build RAG, agents and evals by hand, and essays on the systems around them.
Build a working RAG system by hand, one concept at a time, with runnable code and interactive figures.
Grow one agent from a bare loop into a durable, observable, secured production system.
Build the measurement stack by hand, one concept at a time, so you can tell a real 2% gain from noise.
25 min read Available in EN · TR
Training an LLM happens once. Inference happens every time someone hits enter. These are my notes on what it actually costs to serve an open-source model: where the GPU memory goes, why the KV cache and not the weights is the thing that bites you, what happens inside a single forward pass, and how quantization buys back 3x the throughput for less than a point of accuracy. Part 1 of a walkthrough of efficient LLM serving.
5 min read Available in EN · TR
Chunk size, embeddings, re-rankers: the usual suspects. But the language of your corpus quietly shapes every layer of the pipeline, and reasoning models make it decisive.
10 min read Available in EN · TR
The EU's aviation regulator just published a 239-page concept paper on making AI safe enough to fly. From artificial narrow intelligence to learning assurance and the W-shape process, the mental model behind trustworthy aviation AI, explained for non-experts.
Production LLM systems · Computer vision · Research · Edge AI
Projects with a short write-up each. The full list is on the projects page.
RAG · Knowledge graph · Azure · Production
A production RAG chatbot for families in Italy living with autism, built for Fondazione Paideia.
Copilot Studio · Azure AI Search · WCAG · Evaluation
Two live Copilot Studio chatbots for an international accessibility organization, with accessible chat pages, a search pipeline and an evaluation set.
Dataverse · Workflow · AI drafting · Power Platform
An editorial workspace on Microsoft Dataverse that takes factsheets from first draft to publication, with AI help, checked status changes and an outside review.
ALPR · OCR · FastAPI · PyPI
License-plate recognition from detection to privacy-aware storage, packaged on PyPI and served over a FastAPI HTTP and WebSocket API.
YOLOv8 · ByteTrack · TensorRT · Jetson Nano
A pool-safety detector that tells adults from children with a custom YOLOv8m model, tracks each person and warns when a child is alone.
RAG · GraphRAG · Agents · Evaluation
A controlled comparison of five RAG pipelines for multi-hop requirements traceability, across 13,456 runs and four robustness axes.
Video diffusion · LoRA · arXiv · Single GPU
A two-stage pipeline that adapts Wan2.1 I2V-14B with LoRA on a few short clips and generates 720p scenes in the look of a historical drama.
Speech · Local LLM · Jetson Orin · Safety
A voice assistant that turns spoken Turkish or English commands into actions for an operator console, running offline on a Jetson Orin.
Production RAG · Generative video · Vision inspection · Automation
From vision inspection on manufacturing lines to production RAG chatbots.
Aug 2025 – Feb 2026
Built Odigos, a production RAG chatbot for autism spectrum families in Italy (Fondazione Paideia). End-to-end pipeline: knowledge graph traversal + vector search + semantic reranking with Cohere and GPT-4.1. Deployed on Azure Container Apps with Neo4j, pgvector, CI/CD via GitHub Actions.
Aug 2025 – Feb 2026
Fine-tuned Stable Diffusion (WAN 2.1) for Img2Vid pipelines. Built Graph-RAG and LightRAG systems with Neo4j. Managed Azure deployments and built n8n-based agent workflows for unstructured data extraction.
Oct 2022 – Sep 2025
Built YOLO + PyTorch inspection pipelines that cut destructive testing from 9/day to 3/day. Automated welding parameter optimization with XGBoost. 5000 hrs/yr productivity gain via RPA.
Apr 2022 – Oct 2022
Implemented RTLS (UWB, Bluetooth) for asset tracking. Built operator tracking with Raspberry Pi 4, YOLO, TensorFlow, and MediaPipe. Predictive maintenance models with XGBoost and LightGBM.
Jan 2022 – Apr 2022
Built an OCR-based product recognition system integrating AI/ML with PLC (Snap7) for production line automation.
Mar 2025 – Present
M.Sc. AI Engineering for Autonomous Systems, Technische Hochschule Ingolstadt (THI)
2018 – 2023
B.Sc. Electrical and Electronics Engineering, Uludag University (GPA 3.48 / 4.0)
2021 – 2022
Erasmus+ Exchange Program, University of Zilina (GPA 3.88 / 4.0)
Video generation · RAG · Datasets · Object detection
Papers and preprints, 2023–2025.
2025
arXiv:2510.27364
M. Akarsu, K. Catay, S. B. Vedat, E. K. Yarkan, I. Senturk, A. Sar, D. Eksioglu
2025
arXiv:2511.16700
M. Akarsu
2025
arXiv:2512.18748
R. K. Karaman, M. Akarsu
2023
IEEE INISTA 2023
Z. E. Kaymakci, M. Akarsu, C. N. Ozturk