AI R&D Engineer – LLM, Prompt & Agent Optimization • On-site in Izmir

Arfitect; Konak; On-site; TRY 350,000 - 550,000; Full time

Job summary

Arfitect in Izmir, Turkey, invites an engineer to join its AI R&D track. You will own experiments, evaluations, prompts and agent designs, collaborating with the current AI R&D and applied AI engineers.

This hands-on role focuses on making AI systems smarter, cheaper and more reliable, rather than client delivery deadlines. You will design benchmarks, tune prompts, explore multi-agent coordination, and evaluate safeguards.

Qualifications

• At least 1 year of hands-on work with LLMs – prompt engineering, agent development or LLM evaluation – in professional, academic or serious personal projects.
• Solid Python and the engineering hygiene to keep experiments reproducible: Git, turning notebooks into modules, clean code.
• Deep practical understanding of how LLMs behave: context windows, tool calling, structured outputs, sampling, common failure modes.
• Experience with multiple LLM providers and open-source models, and a practical grasp of how they differ in capabilities, pricing, latency, context handling and tool-calling behaviour.
• Experience with an agent framework (LangGraph, LangChain, LlamaIndex, Semantic Kernel or similar).
• Hands-on RAG experience and familiarity with vector stores.
• Strong research skills: you can work through AI papers, technical reports, model architectures and implementation details, and turn relevant findings into practical experiments.
• An experimental mindset: you form hypotheses, measure, and change your mind when the data says so.
• Ability to explain findings clearly to engineers and to non-technical stakeholders.

Responsibilities

• Track, test and benchmark new models, techniques and tooling as they come out, and give the team a clear verdict on what matters and what doesn't.
• Design, optimize and maintain our prompts and agent instructions: tool use, planning, memory, multi-agent coordination, failure recovery.
• Build evaluation sets and experiment harnesses that measure quality, robustness, latency and cost, and use them to iterate.
• Run experiments on retrieval quality: chunking, embeddings, hybrid search, reranking, context strategies.
• Design and evaluate safeguards for our prompts and agents – prompt-injection defences, input/output validation, guardrails – and measure their effect on security and reliability.
• Fine-tune and evaluate open-source models (LoRA/QLoRA) and test local/on-prem deployment (vLLM, Ollama) for clients with data-residency needs.
• Turn ideas into working prototypes and PoCs; hand validated ones to our applied AI engineers for production, and take their real-world failure cases back into your work.
• Codify what you learn as reusable assets: prompt libraries, agent skills, evaluation templates, internal guides and demos.

Skills

LLM prompt engineering

Python

Experimentation

LangChain

LangGraph

RAG

Vector stores

Communication

Education

Bachelor's Degree in CS/Engineering

Tools

vLLM

Ollama

LangChain

Semantic Kernel

Job description

Full-Time

About Arfitect

Arfitect, founded in 2017 and based at Ege Teknopark, Izmir, is an AI-first custom software and AI consultancy. We build AI-powered enterprise applications for leading companies in energy, manufacturing, aviation and retail, and we develop our own AI SaaS products. Our culture values teamwork, innovation and continuous learning – a place where talent thrives.

Your Role

We are growing our AI R&D track and looking for an engineer whose job is to find out what actually works – and to make our AI systems smarter, cheaper and more reliable. You will work alongside our current AI R&D engineer and our applied AI engineers. This is a hands-on research and development role: you own experiments, evaluations, and the prompts and agent designs the rest of the team builds on, rather than client delivery deadlines.

Responsibilities

• Track, test and benchmark new models, techniques and tooling as they come out, and give the team a clear verdict on what matters and what doesn't.
• Design, optimize and maintain our prompts and agent instructions: tool use, planning, memory, multi-agent coordination, failure recovery.
• Build evaluation sets and experiment harnesses that measure quality, robustness, latency and cost, and use them to iterate.
• Run experiments on retrieval quality: chunking, embeddings, hybrid search, reranking, context strategies.
• Design and evaluate safeguards for our prompts and agents – prompt-injection defences, input/output validation, guardrails – and measure their effect on security and reliability.
• Fine-tune and evaluate open-source models (LoRA/QLoRA) and test local/on-prem deployment (vLLM, Ollama) for clients with data-residency needs.
• Turn ideas into working prototypes and PoCs; hand validated ones to our applied AI engineers for production, and take their real-world failure cases back into your work.
• Codify what you learn as reusable assets: prompt libraries, agent skills, evaluation templates, internal guides and demos.

Requirements

• At least 1 year of hands-on work with LLMs – prompt engineering, agent development or LLM evaluation – in professional, academic or serious personal projects.
• Solid Python and the engineering hygiene to keep experiments reproducible: Git, turning notebooks into modules, clean code.
• Deep practical understanding of how LLMs behave: context windows, tool calling, structured outputs, sampling, common failure modes.
• Experience with multiple LLM providers and open-source models, and a practical grasp of how they differ in capabilities, pricing, latency, context handling and tool-calling behaviour.
• Experience with an agent framework (LangGraph, LangChain, LlamaIndex, Semantic Kernel or similar).
• Hands-on RAG experience and familiarity with vector stores.
• Strong research skills: you can work through AI papers, technical reports, model architectures and implementation details, and turn relevant findings into practical experiments.
• An experimental mindset: you form hypotheses, measure, and change your mind when the data says so.
• Ability to explain findings clearly to engineers and to non-technical stakeholders.

Who You Are

• Curious to the point of obsession about how models and agents behave.
• Rigorous about measurement: you trust numbers over impressions.
• Generous with what you learn – you make the whole team better.
• Comfortable in a field that changes every month.

Nice to Have

• Fine-tuning experience (Hugging Face, PEFT).
• On-prem LLM serving and inference optimization.
• LLM observability and evaluation tools (Langfuse, LangSmith, Ragas or similar).
• Speech and voice models (Whisper, Azure Speech, voice agents).
• Publications, open-source contributions, or a public portfolio of LLM experiments.

Education

Bachelor's Degree in Computer Science, Software/Computer Engineering, Mathematics or other related field from a four-year College or University.