Lead Research Engineer, Search & Retrieval

About The Role

Retrieval is the ceiling on everything above it. An agent working a legal, tax, or regulatory question is only as good as the evidence handed to it, whether it can find the controlling authority in a corpus of millions of documents, weigh sources that conflict, and be honest about what it doesn't have. Every higher-order capability we ship depends on retrieval being trustworthy first.

Retrieval is the ceiling on everything above it. An agent working a legal, tax, or regulatory question is only as good as the evidence handed to it, whether it can find the controlling authority in a corpus of millions of documents, weigh sources that conflict, and be honest about what it doesn't have. Every higher-order capability we ship depends on retrieval being trustworthy first.

This role, in TR Labs, owns the engineering behind that layer: next-generation search and retrieval serving both traditional search experiences and agentic AI workflows over large collections of legal, tax, and regulatory content. Multiple product teams depend on what it delivers.

What makes this research engineering rather than software engineering is that the answer isn't known when you start. Whether a different ranking model, a hybrid retrieval strategy, a new chunking scheme, or an agentic retrieval loop actually makes results better is an empirical question — and an easy one to get wrong, because a metric moving is not the same as retrieval improving.

You form the hypothesis, isolate the variable, read the numbers honestly, and kill the idea when the data says to. Then you do the part many researchers don't: make the winning version production-grade, ship it, and keep it healthy.

You will work shoulder-to-shoulder with applied scientists, building on their models and research directions and feeding production evidence back into the science. As a Lead you own end-to-end delivery, you deliver through the people around you, and you are the person we rely on to know the details, look around corners, and tell us early when something is going sideways.

You are unusually rigorous with evidence. You reach for a baseline, an ablation, and a control before you trust a result, and you have the taste to know which experiments are worth running and which are not.

You have launched search systems, not just built them, but operated them, scaled them, debugged them at 2am, and measured whether they actually made retrieval better.

You build with AI tooling rather than around it, and you bring the same skepticism to what a coding agent hands you as to what an experiment tells you.

You don't wait to be handed a problem. Given a messy project, you can work out what the most impactful next thing to do is and go do it.

You can explain your work to engineers, scientists, and product stakeholders alike: defend a design choice, and update on evidence when someone shows you a better one.

• Own end-to-end delivery of significant search and retrieval projects, accountable for the outcome, the quality and timeline, and the system once it is live
• Act as technical lead for a squad of 3–5 engineers: set direction, break down the work, review designs and code, and unblock the team
• Partner closely with applied scientists, build on their models, ranking approaches, and research directions, and feed production evidence back into the science
• Run the exploration → POC → proof of value → productionization loop, and decide what to try next, including what not to try
• Design and build retrieval architectures, ingestion and indexing pipelines, and ranking and re-ranking systems on Open Search and Vespa
• Build the retrieval infrastructure that agentic AI workflows depend on, and the search agents themselves: tool-facing retrieval APIs, agentic query planning and multi-step retrieval, RAG pipelines, hybrid and semantic retrieval, and query understanding
• Build evaluation that actually discriminates — offline relevance harnesses, golden and labeled sets, online A/B tests, and end-to-end agent quality measurement designed to separate real improvement from a number that happened to move, and to keep discriminating as the models get…