Lead AI Engineer

Coupa is the platform companies run their spending on - sourcing procurement invoices payments suppliers contracts. It is the system of record for how large organisations decide what to buy from whom and at what price.

That makes it one of the most unusual datasets in enterprise software: a single network of 10M buyers and suppliers and $10 trillion of transacted spend to date - quotes bids awards orders invoices contracts and the documents behind every one of them. Multimodal longitudinal and tied to outcomes measured in real money.

Rossum joined Coupa earlier this year. We brought the document understanding layer - our proprietary T-LLM (transactional LLM) architectures which we design and train from scratch and which read the worlds messiest business documents in production millions of them every week. Now we are pointing the same in-house research capability at a much bigger problem: not just reading the documents but acting on them.

We are expanding our Data Capture Research team in Prague with a Senior Data Scientist to work on Sourcing.

Which suppliers get invited. How the event is structured. How bids that differ in price lead time quality risk and carbon get compared at all. What a fair price even is. When to award to whom and how to split the award across suppliers.

For a researcher this is unusually open ground - not one model family but several on the same data:

• Recommendation and retrieval. Supplier discovery: matching demand to the right suppliers across a 10M-node network.

• Forecasting and should-cost modelling. What this category in this region at this volume should cost right now.

• Game theory and mechanism design. Auction formats bidding behaviour incentives and competitive dynamics between real counterparties.

• Combinatorial optimisation. Award allocation under volume capacity and multi-sourcing constraints.

• Multimodal document understanding. RFPs specs quotes and contracts carry the actual requirements - and our T-LLM foundations already give us a head start there.

Part of the work is choosing the right instrument for each - neural networks gradient boosting optimisation bandits mechanism design - instead of forcing one.

Very little of this has been built with modern ML yet. That is the point of the role: real greenfield problems a dataset nobody else has and a product that ships to companies whose margins depend on getting these decisions right.

You will work in a small senior team of researchers and engineers - the group that built Rossums production models from scratch - with direct access to Product and the AI Platform team. Ideas that work do not sit on paper; they roll into systems used at scale.

• Own sourcing research initiatives end to end - from framing the problem on real spend data through experiments to a model running in production.

• Build what does not exist yet. Most of these problems have no baseline to beat and no off-the-shelf answer - you decide what the first version looks like.

• Turn the network into a training set. Define the datasets labels and benchmarks that make sourcing problems learnable at all: deriving supervision from historical events and their outcomes and building evaluation the team can trust.

• Design and run bold experiments with a hacker mindset - fast prototypes honest baselines and offline evaluation that actually predicts online behaviour.

• Build on our T-LLM foundations wherever documents carry the signal - RFPs specs quotes contracts - reusing in-house architectures we train ourselves.

• Ship with deployment in mind: inference cost latency robustness and what happens when a recommendation is wrong in front of a buyer.

• Work across the company - Coupas Sourcing product and data teams and our AI Platform team. Document decisions and make the people around you better.

Sourcing needs several kinds of modelling so we are deliberately open about which one you bring.

• Technical leadership. You have set the direction for a research area others worked in - scoping the problems mentoring the people solving them and staying accountable for whether the models actually shipped.

• 10 years in applied ML data science ML engineering or quantitative research with models or decision systems you took into production.

• Strong Python and real comfort with messy large-scale data - including SQL and the unglamorous work of making a dataset trustworthy.

• Depth in at least one modelling discipline curiosity about the rest. Deep learning recommendation and ranking forecasting and time series optimisation and operations research causal inference and econometrics RL and bandits market and mechanism design or LLM-based systems. Sourcing touches most of these; nobody arrives holding all of them.

• Scientific rigour. Strong experiment design healthy scepticism about your own metrics and real care about leakage baselines and evaluation that survives contact with production.

• Ownership and curiosity. You are comfortable in a greenfield ambiguous problem space and you will talk to product people and procurement experts to find where the value actually is.

• Interest in procurement supply chains or market design is welcome but not required - we will teach the domain.

• Greenfield problems in a mature product: Modern ML has barely been applied to sourcing inside a platform that already has the users the workflows and the data.

• Data nobody else has: $10 trillion of transacted spend 10M buyers and suppliers and the documents behind all of it.

• We train our own models: Proprietary T-LLM architectures designed and trained in-house - not a wrapper around someone elses API.

• Real ownership short path to customers: You frame the problem choose the method and see it working in front of buyers - no research-to-product handoff.

• Global impact: Technology used every day by companies around the world.

• Experiment-driven culture: Pragmatic delivery and quarterly recognition for standout research contributions.

• Compute and tools: Frontier LLMs on tap for your own work and our high-end GPU and large-memory clusters to train on.

• 33 days off: PTO personal days your birthday and two company wellness days. Parental leave on top.

• Prague Karlín: Inspiring workspace and full tech setup including a 200 m² terrace with views of Prague Castle.

Required Experience:

IC