The Position

MetaDefender Core is OPSWAT's file security platform, trusted where the stakes are highest: nuclear plants, defense networks, and the banks that move the world's money. The most sensitive of those customers cannot send a single byte to anyone's cloud, so the AI must run entirely on their hardware, inside the product. 

This role is for the engineer who would rather own a model than call someone else: you build the training data, the fine-tuning, the evaluation, and you make the call on what ships. Every AI assistant you have ever used lives in a data center; yours will run on the customer's own hardware, inside some of the most protected networks in the world. 

What You Will be Doing 

• Own the model track end to end: training data, fine-tuning runs, evaluation, and release, with scope and ownership growing by level. 
• Build and maintain the evaluation gate for every model to change and make it something the team trusts more than its own intuition. 
• Build the data pipelines that produce training datasets, enforcing grounding so the model can only state what it has actually read. 
• Shape how the agent behaves: its instructions, its tool contracts, and how it recovers when the model gets something wrong. 
• Own the model's security posture: harden it against prompt injection and misuse, make sure it never exposes internal information, and never lets itself be steered into guiding users toward unsafe actions. 
• Serve models on constrained customer hardware: quantization, inference runtimes, and the memory / latency / quality trade-off. 
• Ship into a production Rust service, collaborating closely with the engineering team, Product Management, and QA. 

What We Need from You 

• Bachelor's degree in a technical field or equivalent practical experience. 
• Experience fine-tuning open-weight LLMs (LoRA/QLoRA or similar). 
• Strong Python, including building data pipelines for training datasets. 
• Experience keeping model output grounded: retrieval-backed generation, hallucination control, and output traceable to a source the model reads. 
• Experience with LLM serving and inference (e.g. vLLM, Ollama), including quantization and the memory / latency / quality trade-off on constrained hardware. 
• Familiarity with tool-calling / agent-based LLM applications. 
• Ability to design and run model evaluations for non-deterministic systems: you can name a metric you designed and the number you moved. 
• Working knowledge of SQL. 
• Strong verbal and written communication skills in English: the language of the team, the docs, and the prompts. 
• Self-motivated, adaptable, and comfortable owning a track in a fast-paced, team-oriented environment. 

It Would be Nice if You Had 

• Rust literacy (able to read and modify existing service code). 
• Experience with agent frameworks (e.g. ADK-style). 
• Awareness of LLM security risks: prompt injection, data leakage, misuse. 
• Compact models deployed on-premises, offline, or resource-constrained environments. 
• Evaluation tooling for LLM systems: LLM-as-judge, regression suites, human review loops. 
• Cybersecurity concepts, or experience building components of a security product.