AI Research Engineer

About the Role
Research Engineers here are researchers first — deep in AI and machine learning, with the mathematical and scientific grounding to push the frontier of what our agents can do. But you don't stop at the paper. You're an experimentalist who closes the loop, taking your own research and implementing it into production software systems that finance teams and auditors can trust. You view research and engineering as two sides of the same coin.

This is a founding role on what is, in effect, our clients research lab — a NeoLab for finance. You won't simply execute a roadmap handed to you. You'll be a genuine thought partner to the founders: helping decide where this field is heading, not just two months out but one to two years ahead, and where we should place our research bets to stay at the frontier. You'll own a research agenda end-to-end and shape both the technical direction and the team we build around it.

What You'll Do
Shape the frontier with the founders. Work shoulder-to-shoulder with the founders on where finance AI is heading — one to two years out, not just the next sprint — and turn that thesis into the research agenda we commit to.
Make correctness provable. Bring formal-verification methods — tools like Lean 4 and Catala — to bear on accounting logic, working toward a formally verifiable treatment of GAAP so an agent's conclusions can be checked against the standard, not merely trusted.
Make agents trustworthy. Develop evaluation frameworks and synthetic data generators that hold agents to an audit-grade bar: groundedness, correctness, and knowing when to defer to a human.
Bridge research and production. Take ideas from experiment to clean, deployed code, building the training and inference infrastructure to support them.
Build the core intelligence. Design, train, and fine-tune LLMs and agentic systems for accuracy, reliability, and efficiency on domain-specific finance tasks.
Scale the systems. Architect robust pipelines and distributed infrastructure that reliably support large-scale, real-world operational demand.
Own a research agenda. Choose impactful problems, autonomously carry out long-running projects, and translate frontier research into shipped features and, where it fits, published work.

Sample Projects
Building a synthetic data pipeline that generates realistic, messy financial scenarios to train and stress-test agents.
Designing an evaluation harness that measures whether an agent's reconciliation output is correct and audit-defensible.
Fine-tuning and distilling a model to run accurately under production latency and cost constraints.
Researching context-engineering methods that let an agent reason over transactions, ledgers, and contracts without bloat, drift, or data leakage.
Designing verification and guardrail methods so agents escalate uncertainty instead of guessing — and improve measurably from human corrections.
Building a formally verifiable representation of GAAP — using a language like Lean 4 or Catala — so an agent's accounting conclusions can be mechanically checked against the standard rather than taken on trust.

Basic Qualifications
PhD in Computer Science, Machine Learning, AI, NLP, or a related field (degree obtained on or before start date).
2+ years of experience as a research engineer, ideally at a startup — though we'll waive this for exceptional candidates, including those coming straight out of a PhD.
A strong publication record in top AI venues (NeurIPS, ICML, ICLR, ACL, EMNLP, or equivalent) on topics across AI agents, LLMs, or ML.
Hands-on experience training and fine-tuning LLMs, with deep proficiency in Python and PyTorch.
Strong software engineering skills with a proven track record of building complex systems.

Preferred Qualifications
Publications or production experience in LLM agents — tool use, planning, multi-step reasoning, or orchestration.
PhD focus on NLP, or equivalent industrial NLP research experience, with hands-on instruction-tuning and model adaptation.
Experience deploying a fine-tuned LLM to production, including RL/RLHF, DPO, or other policy optimization methods.
Training and serving models under real-world latency and cost constraints — quantization, distillation, sparsification, or compression.
Designing synthetic data generators, RAG pipelines, or rigorous evaluation methodology for unstructured, high-stakes tasks.
Tech stack
Python, PyTorch, LLMs, Lean 4, Catala, RAG, RLHF, DPO