Most finance chatbots are a language model holding a summary of your spending. Ours sits on top of a ledger that can answer exactly, so the interesting engineering is not fluency — it is grounding, tool design, and the policy layer that decides when the assistant may act on its own and when it must ask.
You would own the assistant end to end: the tool loop, the categorisation models, the evaluation harness that tells us a change made it better, and the guardrails that keep it from moving someone's money on a hunch.
What you would do
- Design the tool surface the model acts through, and the permission policy behind each tool.
- Build evaluations that are honest — regression sets from real failures, not vibes.
- Improve transaction categorisation and counterparty resolution with a measurable confidence signal.
- Keep latency and cost inside a budget the product can afford at scale.
What we look for
- Shipped an LLM-backed feature that real users depended on.
- Strong Python; comfortable with retrieval, tool calling and streaming.
- You are rigorous about evaluation, and sceptical of your own demos.
Nice to have, never required
- Experience with agentic tool loops and permissioned actions.
- Classical ML for classification problems, not only prompting.
If you are on the fence
Every list above describes the person we imagine, not a checklist we score. People who meet most of it and want the rest have done extremely well here. If the role reads like the job you want next, that is enough reason to write to us — tell us which part you would have to grow into and we will be straight with you about whether it works.
Send whatever shows your work best — a CV, a repository, a portfolio, or a few paragraphs. A hiring manager reads every one, and you will hear back within two business days.
Apply by emailcareers@lumaledger.app