Agentic and generative features

• Build the LLM-backed features in the product, including the guided-question loops in the BD Centre and the generative surfaces around dossiers, feeds and alerts.
• Design agent and tool-use flows that fail safely, degrade to something useful and never invent a client, a deal or a quote.
• Own prompt architecture as engineering rather than as text: versioned, tested, reviewable and cheap to change.

Evaluation and quality

• Build the eval harness. Define what good looks like for each AI surface, build the datasets, and make regression visible before a release rather than after it.
• Instrument quality in production: hallucination rate, retrieval hit rate, refusal behaviour, latency and cost per interaction.
• Run the experiments that decide model choice, and be willing to conclude that a smaller or cheaper model is the right answer.

Data and extraction

• Work with the data pipeline that ingests legal news and deal coverage, and own the extraction that turns it into structured, queryable records.
• Improve precision and recall on that extraction over time, and know the current numbers for both.
• Feed clean data into the rankings engine, and understand enough of its methodology to spot when the inputs are wrong.

Engineering and platform

• Ship production code into a TypeScript and Python stack running on AWS, and take responsibility for it in production.
• Own cost, latency and reliability of the AI layer, including caching, batching, fallbacks and rate limits.
• Build the internal tooling that lets non-engineers inspect, correct and improve model output without asking you.

Judgement and influence

• Tell Product what is feasible, what is expensive and what is a research project rather than a sprint.
• Push back where a feature is being specified as an AI feature when it should not be one.
• Set the standard and the practices that the next AI hires will work to.

WHAT WE ARE LOOKING FOR

• 5-8 years in software engineering, with at least two spent building LLM-backed systems that real users depend on. Production experience, not prototypes.
• Deep applied retrieval experience: you have built RAG systems that worked, and you can explain in detail why the first version did not.
• Rigour about evaluation. You have built eval infrastructure and you treat an unmeasured AI feature as an unfinished one.
• Strong engineering fundamentals in Python and TypeScript, comfortable in AWS and in production systems rather than notebooks.
• Fluency across the current model landscape and the tooling around it, with the judgement to pick the boring option when it wins.
• Comfort with ambiguity and with owning a domain alone. You will not have a team to delegate to for some time.
• Directness. You raise problems early, you argue your case on evidence, and you change your mind when the data says so.

PREFERRED

• Experience in legal, financial or another domain where accuracy is a professional obligation rather than a nice-to-have.
• Experience with entity resolution, knowledge graphs or document-heavy extraction at scale.
• Fine-tuning, distillation or model serving where it earned its keep against a hosted API.
• Early-stage or pre-launch experience, and a working understanding of what unit economics mean for an AI product.
• An open-source record, a paper, a side project, or anything else that shows us what you build when nobody is assigning it.

Rainmaker is an equal opportunity employer. We are committed to diversity and inclusivity.