Senior RAG Engineer

About Newcode.ai

Newcode.ai is a fast-growing legal tech and agentic AI company transforming how legal work is done. With teams across Norway, Sweden, Ireland, the US, and growing, we work at the intersection of law, technology, and intelligence. We move fast, think big, and take pride in doing things the right way.

Requirements

• Own the retrieval pipeline end to end: document parsing and chunking, embeddings, indexing, and keeping all of it current as clients add and change material.
• Exposed through FastAPI, with ingestion and reindexing running as background jobs.
• Design how we search. Combining semantic and keyword retrieval in Qdrant, fusing ranked lists, filtering on metadata, and reranking so the handful of results we pass to the model are the right ones.
• Build agentic retrieval: query decomposition, the tools a model uses to search and navigate documents, multi-step loops that know when to stop, and the cost and latency budgets that keep them honest.
• Build the evaluation layer that tells us any of this is working: golden sets, retrieval metrics, regression tests on realistic client data, and tracing good enough that a bad answer leads you back to the chunk that caused it.
• Ship it as fast, dependable services in Python and FastAPI, with the heavy work - ingestion, embedding, reindexing - running as background jobs that hold up under load.
• Take data security and isolation seriously. Clients hand us privileged material, and tenants stay strictly separated.
• Make retrieval hold up across our clients' languages as well as it does in English.

Qualifications

• At least five years building backend systems that run in production, and at least two of them shipping retrieval or RAG systems real users depend on.
• Backend depth without production retrieval won't be enough for this role, but our Senior Backend Engineer role might be the better fit.
• You've diagnosed a retrieval regression in production and fixed it. You can tell us what broke, how you found it, and what the numbers were before and after.
• You've built and maintained a golden set. How many queries, who labelled them, which metrics you trust, and a change you shipped or killed because of what they told you. You've run a vector index in production: picked index parameters, dealt with the memory and latency trade-offs, and reindexed without taking search down.
• Strong Python. Comfortable reasoning about FastAPI or a close equivalent, PostgreSQL and Redis under load.
• You don't ship a retrieval change because the output looked better on the three queries you tried by hand. You own things end to end — including the boring maintenance and the tech debt nobody assigned you — rather than building the interesting part and handing off the rest.
• You track what's moving in vector databases and LLMs because you're curious, not because it's the job. Show us the side project, the benchmark you ran for fun, or the repo where you tried something before it showed up in everyone else's stack.
• You've worked with AI coding assistants and have a view on where they help and where they don't. If your current employer forbids them, that's not a mark against you. Tell us how you'd review AI-written code instead.
• You're fine in a startup that changes direction. Decisions get made, then revisited.