Member of the Technical Staff — AI/ML

What matters most

• Production-grade financial agents

  Build tool-using LLM agents that surface insights, recommend next steps and execute approved tasks. The posting emphasizes deployed systems and understanding failures in tool use, planning, state, recovery and human oversight.

• Useful workflow improvements

  Refine invoice matching, payment reconciliation and financial document processing as features integrated with Stuut’s platform, rather than standalone prototypes.

• Practical model and pipeline choices

  Apply LLMs and RAG to financial workflows, fine-tune on proprietary data when useful, and build reliable, maintainable ingestion-to-inference pipelines with cost-conscious model routing. The role also calls for experience deploying classical ML models.

• Evaluation and continuous improvement

  Establish golden datasets, agent tracing, regression tests on pull requests and A/B tests. Instrument deployed features, monitor outcomes and iterate to catch silent regressions.

• Translate needs into customer impact

  Work with Product, Engineering, Data and customers to turn financial-operations needs into technical solutions with measurable business value.

Likely challenges to think through

• Balancing agent action with oversight

  Likely preparation challenge, not a confirmed problem: executing approved financial tasks can be more consequential than recommending them. Prepare to explain how an agent would track state, detect tool or planning failures, recover safely and involve a human when needed.

• Choosing capability without unnecessary cost

  Likely preparation challenge: better model output may increase inference cost or pipeline complexity. Prepare a decision framework for when to use RAG, fine-tuning, model routing or classical ML, and how to compare quality, reliability, cost and maintainability.

• Connecting test results to real outcomes

  Likely preparation challenge: offline evaluations may miss workflow edge cases or fail to predict customer impact. Prepare to describe how golden datasets, traces and pull-request regressions would complement online monitoring and statistically sound A/B tests.

Topics to prepare

• Agent reliability and human oversight

  The posting emphasizes tool use, planning, state, recovery and human-in-the-loop failure modes. Revision exercise: walk through an invoice-matching agent that recommends an action but may execute it only after approval; identify how it handles a failed tool call, stale state and an ambiguous match.

• RAG versus LLM fine-tuning

  Stuut calls for RAG in document- or data-intensive financial workflows and fine-tuning proprietary data where useful. Revision exercise: compare retrieval, fine-tuning and a simpler baseline for answering questions from financial documents; state what evidence would justify each choice.

• Classical ML model decisions

  The posting explicitly names deployed risk-scoring, forecasting or ranking models, including feature engineering, selection, evaluation and calibration. Revision exercise: outline a ranking model for payment follow-up, then explain how you would check both ranking quality and score calibration.

• Offline and online evaluation

  Golden datasets, evaluations in CI, A/B testing and statistical significance are explicit priorities. Revision exercise: design an offline regression check and an online comparison for invoice matching; define what would count as a meaningful improvement or a silent regression.

• Production pipeline tradeoffs

  The role covers ingestion through inference, reliability, maintainability, cost efficiency and model routing. Revision exercise: sketch that pipeline for financial document processing and explain where a routing decision could reduce cost without compromising an evaluation target.

• Receivables workflow context

  The role names invoice matching and payment reconciliation within broader B2B order-to-cash workflows; AR/AP experience is a bonus, not a stated requirement. As preparation, trace a hypothetical invoice and payment through matching, reconciliation and exception handling, marking where an AI recommendation would need review.

Skills to revise

• Implement an approval-gated agent

  The posting calls for production-ready, tool-using LLM agents that execute approved tasks. Practice in Python with a small synthetic workflow: return a proposed reconciliation action, require explicit approval before execution, and test failed calls and recovery. This is a suggested exercise, not a specified employer assignment.

• Build and assess a document RAG pipeline

  Stuut seeks LLM integration and RAG for financial workflows. As practice, ingest sample financial documents, retrieve evidence for an invoice question, and inspect incorrect or unsupported answers. Use sample data rather than assuming access to Stuut’s proprietary data.

• Train and calibrate a classical model

  The posting asks for hands-on feature engineering, model evaluation and calibration. Practice with synthetic tabular payment data in Python and a listed tool such as scikit-learn: train a risk or ranking baseline, check held-out performance, and examine calibration.

• Create regression data and traces

  Golden datasets, agent tracing and regression-on-PR are named responsibilities. Practice assembling a small labeled set of matching cases, recording each agent’s tool calls and outcomes, and writing a test that fails when a known case regresses; the exercise is illustrative, not a claim about Stuut’s test setup.

• Instrument a controlled rollout

  The posting calls for monitoring outcomes, A/B testing and model and prompt versioning. Practice defining version identifiers, outcome metrics and a comparison plan for two invoice-matching variants, including how you would detect degradation after deployment.

• Map the financial workflow

  Sketch how invoice matching or payment reconciliation fits into Stuut’s order-to-cash work. Use the posting’s scope as context; mark any assumptions you would need to validate.

• Prepare a production project walkthrough

  If you have a relevant shipped AI/ML product, bring a concise account of the business problem, your decisions, deployment, measured outcomes and lessons learned. Distinguish production evidence from prototype work.

• Practice a tool-using agent design

  As an interview exercise, diagram an agent that recommends a reconciliation step and executes it only after approval. Explain tool-use, planning and state failures, recovery paths and human oversight.

• Compare LLM pipeline choices

  Draft an ingestion-to-inference design for financial documents. Show where RAG, fine-tuning or model routing might help, and how you would test quality, reliability and cost before choosing.

• Assemble classical ML evidence

  If applicable, prepare one risk-scoring, forecasting or ranking example covering feature engineering, model selection, evaluation, calibration and deployment. The posting seeks deployed classical ML as well as LLM experience.

• Build an evaluation and monitoring plan

  Outline a golden dataset, agent traces, pull-request regression checks and an online A/B test. Specify failure metrics, a statistical decision rule, and how model and prompt versions would be monitored.

• Write collaboration questions

  Prepare questions for Product, Engineering, Data and customer-facing partners about workflow pain points, approval boundaries, available evaluation data and how they measure customer impact.