I build AI systems that survive production.

• 

  Document Intelligence Suite

  Three parsers reading invoices, emails and timesheets past 90% field accuracy, including the ones that arrive as photos of paper.

  2026LLMsRAGDocument AI
• 

  Sales Comment Moderation

  A locally trained LLM that moderates comments on sales listings in-house, so customer text never leaves the building to get a label.

  2026LLMsFine-tuningContent Moderation
• 

  Sales Lead Proposal System

  A lead qualification and proposal engine that predicts purchase intent with semantic similarity rating, instead of asking an LLM to score leads directly.

  2025LLMsEmbeddingsSemantic Similarity
• 

  Enterprise Knowledge Graph

  Agentic pipelines that turn scattered public data on European enterprises into a graph you can query.

  2024Agentic AIKnowledge GraphsNLP

I've spent most of my career on the part that comes after the model works. The async services, the retries, the evaluation harness, the deployment. A model that's right once is a demo. Keeping it right on the thousandth request is the actual job, and it's usually where the time goes.

Right now I work mostly on LLM agents: multi-step tool use, retrieval over knowledge graphs, and a lot of unglamorous plumbing to make non-deterministic systems reliable enough to sit inside a business process. Before that I did computer vision on edge hardware for road and rail safety, where a wrong answer at 60 frames per second isn't an abstract cost.

The problems I enjoy tend to have the same shape. Model accuracy is fine; the hard constraint is somewhere else. Latency budgets, inputs nobody cleaned, or the distance between a benchmark score and how the thing behaves on an ordinary Tuesday.

1. AI Solutions Engineer

   Oct 2021 – Sep 2026

   innoscripta SE · Germany · Hybrid

   • Built the agentic extraction pipeline behind a knowledge graph of European enterprise data: LLM agents deciding what to pull, NLP models resolving entities, and the results landing somewhere you can actually query.
   • Led three document-intelligence tools to production (invoice, email and timesheet parsers), all past 90% field-level accuracy on documents that are often just a photo of a piece of paper.
   • Wrapped the agents in async FastAPI services so other teams could integrate over HTTP without touching a model. Docker deployments, CI/CD on AWS.

2. ML Solutions Engineer

   Aug 2018 – Sep 2021

   ISSD Bilişim Elektronik A.Ş. · Turkey · Hybrid

   • Shipped real-time incident detection for tunnels and highways with Intel. Models ran optimized on roadside hardware, since streaming the video to a datacenter and back was never going to be fast enough.
   • Built a 3D container scanner for a port terminal, which raised throughput and cut misreads at the same time.
   • Developed rail-crossing detection with Huawei, flagging vehicles and people in the danger zone while there was still time to act.

• LLM Systems

  • Multi-agent orchestration
  • RAG pipelines
  • Tool calling
  • LangChain
  • Autogen
  • MCP

• ML & Vision

  • NLP
  • Object detection
  • Incident detection
  • Crowd analysis
  • Time series forecasting

• Backend

  • Python (asyncio)
  • FastAPI
  • C++
  • Qt
  • Microservices

• Infrastructure

  • AWS
  • Docker
  • Kubernetes
  • OpenVINO
  • CI/CD

• Data Stores

  • PostgreSQL
  • Qdrant
  • Elasticsearch
  • JanusGraph
  • AWS Neptune
  • Redis
  • MongoDB

1. Incident Detection on Junction Using Image ProcessingMehmet Demirci · arXiv preprint 2104.13437 · 2021
2. Application of the Neural Network Dependability Kit in Real-World EnvironmentsMehmet Demirci · arXiv preprint 2012.09602 · 2020
3. Aiming for Smart Wind EnergyMehmet Demirci · Transactions on Emerging Telecommunications Technologies · 2019