The posting
AI Engineer (Applied LLM Systems)
About Discernis Discernis builds AI driven document intelligence for high stakes legal work. Because our customers handle privileged and regulated matters that most cloud AI cannot touch, we run on premises and in customer controlled environments as well as in the cloud. That makes security, reliability, and reproducibility core product features, not afterthoughts. We work with AmLaw firms and enterprise legal teams where accuracy, explainability, and data control are non negotiable.
The Role You will lead the design and evolution of the AI systems at the core of our product. Day to day, that means inventing new AI powered workflows that help legal teams understand massive document sets, then making them faster and more reliable, and building the evaluations that tell us we are getting it right. You will reach for whatever technique fits the problem, prompt engineering, agentic orchestration, retrieval, or post-training, and because our customers often cannot send data to external providers, much of this runs on models we host and improve ourselves.
What You Will Do
- Design and build new AI powered workflows that solve real problems for legal teams, using whatever gets the best result: agentic and multi step orchestration, prompt engineering, retrieval, and post-training
- Improve the quality, reliability, latency, and cost of existing workflows using that same full toolkit
- Own and develop our evaluation and benchmarking frameworks so we can measure and improve quality, catch regressions, and know we are moving in the right direction
- Design agentic patterns and structured outputs robust enough for complex legal and investigative use cases
- Own self hosted model serving and inference optimization across customer deployments
- Post-train and adapt models where that is the right lever, including fine tuning, preference optimization, and domain adaptation
What You Bring
- Experience building production LLM or ML powered features end to end, with the judgment to know which technique fits which problem
- Hands on experience with agentic or multi step LLM orchestration and tool use
- Breadth across the modern LLM toolkit: prompt engineering and retrieval, and ideally post-training methods such as fine tuning, DPO, or RLHF
- Experience designing and running evaluations for LLM systems
- Hands on experience with self hosted model serving such as vLLM, TGI, or similar
- Strong Python skills and comfort with concurrent or distributed processing
- Bonus: embedding and vector search pipelines, or retrieval augmented generation at scale
- Bonus: background in legal tech, document intelligence, or information retrieval
Tech Environment Python, self hosted inference (vLLM), LLM orchestration and agent frameworks, evaluation tooling.
