AI Data Scientist

About DolfinWe're building the agentic financial ops API for SMB platforms. Fintechs, vertical SaaS, banks, and PSPs use Dolfin to embed AI-native financial ops agents (invoicing, bills, expenses etc) directly into their products. White-label, API-first, production-ready.Our agents are live in production with real SMBs. We're live and moving fast: signed platform clients including a unicorn, pipeline in the $M's, and part of Balderton's latest 'Launched' cohort.We're looking for an AI Data Scientist to own the intelligence layer our agents run on. Every Dolfin agent does the same thing underneath: read a financial document, work out what it relates to, decide whether to act or ask a human.Your job is to make that decision right, at scale, across every platform we're live with.What you'll be doingOwn extraction and matching across AR, AP, expenses and more. Invoices, purchase orders, remittances, bill, receipts, card transactions. Unstructured PDFs, photos, and structured e-invoicing feeds.Own confidence scoring and threshold calibration. Auto-accept vs escalate is the highest-leverage decision in the product. Too loose and we post bad entries to a real ledger. Too tight and the platform's customers do the work themselves.Solve cold start. New platforms go live in days with almost no labelled data for their formats, chart of accounts, or policies. Making that work without a bespoke build per client is an open problem, and it's yours.Build the evaluation infrastructure. Gold sets, regression suites, per-platform accuracy tracking. We find out when something regresses before a client does.Build the correction flywheel. Every fix an SMB makes inside a partner's product is a label. Turn that into measurable accuracy gains.Turn written rules into executable logic. Expense policies and approval rules, interpreted, applied, and explainable.Partner directly with the founders on where the accuracy bar sits and what we ship next.How we buildProduction-first. If it isn't live and reliable, it isn't done.Measured, not intuited. If you can't show the number moved, it didn't.Correctness is non-negotiable. This is someone's ledger. Precision and auditability come first.Pragmatic, not dogmatic. Simple approaches where they work, sophistication only where it earns its keep.AI-native. LLMs as infrastructure, not features.Built to embed. Multi-tenant from the ground up. Nothing we build should overfit to one client.Small team, high trust. Judgement over process, every time.What you'll need3+ years shipping applied ML or AI systems into production, not research or analytics.Hands-on with LLMs in production: structured output, tool calling, retrieval, prompt and pipeline iteration against measured outcomes.Strong production Python. You ship code, not notebooks.You've built evaluation harnesses and treat accuracy as an engineering problem rather than a judgement call.Experience in a domain where being wrong is expensive. Fintech, compliance, legal, insurance, healthcare.Genuinely autonomous: take an ambiguous accuracy problem, make the call, ship the improvement.Comfortable operating without structure, and creating it.Bonus: document AI (OCR, layout parsing, entity extraction), AP/AR or accounting systems knowledge, e-invoicing, or time at an API-led fintech or infrastructure company.Who this is forPeople who want to own the number rather than present it. Builders who've put AI into production where it had to be right. People who make the call rather than wait to be told.Not for: research or PhD-track scientists who want to train models and publish. Analytics data scientists expecting a warehouse, dashboards, and A/B traffic. Fully remote candidates. Anyone who needs tightly scoped tickets.Why usGenuine early-team role with meaningful equity.Real production AI systems with real money on the other side, not experiments.Deep ownership of the layer the entire product depends on, and a hand in setting the accuracy bar from day one.Southbank, London, in-person. Full-time. Let's go!