Artificial Intelligence Engineer

OZSOFT® - Software & Technology Applications; Çankaya; On-site; TRY 350,000 - 700,000; Full time

Benefits offered by this job

Health insurance

Meal allowance

Job summary

OZSOFT® - Software & Technology Applications in Turkey is seeking a machine learning/NLP engineer to build and refine Turkish clinical NLP models and deploy on-premise LLMs with low latency.

You will work with radiologists and HBYS teams to design evaluation pipelines, label data, and integrate systems (REST, Kafka) while upholding strict data privacy and security requirements.

Qualifications

• 3+ years of hands-on ML/NLP experience with Python and PyTorch/Hugging Face.
• Experience deploying LLMs in production with low-latency serving and GPU optimization.
• Strong understanding of evaluation methodology and data privacy.
• Experience with Turkish NLP or motivation to master it.
• Comfort with Linux, Docker and on-prem infrastructure.

Responsibilities

• Build and fine-tune Turkish clinical NLP models (encoder classifiers, NER, negation detection).
• Serve open-weight LLMs on-premise (vLLM / SGLang / TensorRT-LLM) on NVIDIA GPUs, using quantization and latency budgets.
• Design evaluation pipelines with blind gold sets, shadow mode, and inter-annotator agreement.
• Collaborate with radiologists and HBYS team on labeling and integration (REST, CDC / Kafka, read replicas).
• Build RAG pipelines over medical guidelines (Qdrant / Milvus).

Skills

Python

PyTorch

Hugging Face

Turkish NLP

ML deployment

Latency optimization

Tools

Docker

Linux

On-prem infrastructure

REST

Kafka

NVIDIA GPUs

vLLM / TensorRT-LLM

Quantization FP8/FP4

Job description

We are building an enterprise AI platform that runs entirely inside our own data center. No patient data ever leaves the hospital. The platform supports clinicians and finance teams in real time, directly inside our Hospital Information System (HBYS).

What you'll work on

Radiology & clinical report quality control: detect laterality errors, missing findings, measurement and impression inconsistencies, and BI-RADS / TI-RADS / LI-RADS compliance before a report is signed, with sub-second response times.

Clinical decision support: combine radiology, laboratory trends (delta checks), anamnesis and discharge notes into structured, guideline-based action recommendations.

Closed-loop medical CRM: event-driven follow-up that makes sure recommended exams, consultations and surgeries actually happen.

Medical billing intelligence: NLP that matches clinical documentation to billing codes and payer rules, working alongside a deterministic rule engine.

Your responsibilities

Build and fine-tune Turkish clinical NLP models: encoder classifiers, NER, assertion and negation detection, and LLMs.

Serve open-weight LLMs on-premise (vLLM / SGLang / TensorRT-LLM) on NVIDIA GPUs, using quantization (FP8/FP4) and strict latency budgets.

Design evaluation pipelines: blind gold sets, shadow mode, inter-annotator agreement, regression tests.

Work with radiologists and clinicians on labeling, and with the HBYS team on integration (REST, CDC / Kafka, read replicas).

Build RAG pipelines over medical guidelines (Qdrant / Milvus).

What we’re looking for

3+ years of hands-on ML/NLP experience, with strong Python and PyTorch / Hugging Face skills.

Experience deploying LLMs in production. Low-latency serving and GPU optimization are a big plus.

A solid understanding of evaluation methodology. You care about false positives as much as accuracy.

Experience with Turkish NLP, or the motivation to master it.

Comfort with Linux, Docker and on-prem infrastructure.

A strong sense of responsibility for data privacy and security.

Nice to have

Speech technologies (Whisper STT / TTS), OCR and document understanding.

Real clinical impact: your models will run in a live hospital, not a slide deck.

Dedicated on-premise GPU infrastructure and a working prototype to build on.

Close collaboration with physicians and hospital leadership.

[Benefits: salary range, health insurance, meal allowance, etc.]

All applications are treated confidentially.