AI Engineer (LLM / NLP) — Build Our Medical-Video Pipeline

MediBrief turns medical conference lectures into short, faithful animated summaries — the kind conference organizers, medical event companies, and sponsors use to make a 45-minute talk land in three minutes. We've run this as a mostly-manual craft for about two years. We're now building an AI-assisted production system to scale it without losing the thing that makes it trustworthy: fidelity to what the lecturer actually said. We're hiring one mid-level AI engineer to build the front half of that pipeline — the part that takes a raw transcript and produces a fidelity-checked script and a structured, machine-readable spec for the visuals. Why this role is genuinely interesting Our best-practices library is finished work — final scripts, treatments, and videos — with no rulebook written down. The implicit rules live in the examples. The job is to reverse-engineer those patterns and turn them into reliable AI components. If you're the kind of engineer who enjoys inferring the rule from the data rather than being handed a spec, this is your kind of problem. What you'll build Transcript cleaning + a script generator that follows our consistent house structure. A fidelity checker — the heart of the system. It traces every claim in a summary back to the source and flags anything unsupported or drifting (a hedged "may reduce" quietly becoming a flat "reduces" is exactly the kind of thing it must catch). A structured treatment + storyboard spec: machine-readable, per-shot output that the downstream animation step consumes. Evaluation harnesses and iteration loops so quality is measured, not guessed. What we're looking for 2–5 years building real software — someone who ships, not only prototypes in notebooks. Hands-on with modern LLMs: prompting, structured / JSON outputs, retrieval, and evaluation. You don't need to have trained models — you need to make them reliable. Strong at reverse-engineering patterns from examples and working from thin direction. This is the core skill of the role. A clear writer — you'll document what you build and hand parts of it to non-engineer reviewers. Bonus: NLP / information-extraction background; experience with faithfulness, groundedness, or hallucination checks; or any exposure to medical or scientific content. How we work You'll own the front half end-to-end and make the architecture calls yourself, working directly with the founder and the production team whose expertise you'll be encoding. This is a build-it role, not a follow-the-spec one. We build end-to-end first: one lecture all the way through the pipeline, roughly, before deepening any stage. You'll see the whole system fast. Real medical content sits behind human review gates. Your components make those gates fast and reliable; they don't replace the humans. Quick ask before you apply (~10 minutes, no prep needed) We work by reverse-engineering patterns from finished examples, so we'd love a small taste of how you think. Imagine we hand you five pairs of a lecture transcript and its finished summary. In a few paragraphs — or a 2-minute video — tell us: How would you build something that checks the summary says nothing the transcript didn't, and flags anything that drifts? What would you look at first, and how would you know it's working? We're not after a finished design. We want to see how you break the problem down.