Own the models behind ZAAN Contract Review: the systems that read a contract against a firm's playbook, rank every issue by severity and propose a redline a partner would sign off. You will take models from evaluation set to production at the precision legal work demands.

What you will do

• Fine-tune, distil and serve large language models for clause classification, risk scoring and redline generation across 38 languages.
• Build retrieval over playbooks and precedent so every flag cites the clause and the deal it came from.
• Design offline and online evaluation with lawyers in the loop: precision/recall per clause type, citation faithfulness, regression gates in CI.
• Cut latency and cost per document without losing accuracy: batching, quantisation, speculative decoding, caching.
• Work with our security team so customer data never trains shared models and stays in the region the customer chose.

What you bring

• 5+ years building ML systems that run in production, including recent hands-on work with transformer LLMs.
• Strong Python and PyTorch; experience with at least one serving stack (vLLM, TensorRT-LLM, TGI or similar).
• Rigour about evaluation: you design test sets, measure error by category and can explain why a model fails.
• Clear written English. Legal, financial or other high-stakes domain experience is a strong plus.

Nice to have

• Long-context modelling, structured output / constrained decoding, multilingual NLP, especially German, French or Chinese.
• Experience with preference optimisation (DPO, RLHF) on expert feedback.

How we hire

A 10-minute AI interview in English (you are told up front that you are speaking with an AI), a conversation with the team you would join, and a short paid practical task that mirrors real work. We aim to decide within three weeks.