Applied Scientist

Title: AI Scientist/Senior Scientist Company: Stealth Mode AI Company (B Capital Affiliate) About the Company A B Capital affiliate is working to fill several senior, founding-team roles across Dublin, Singapore, and New York. The affiliate is a confidential, stealth-mode AI company building decision intelligence for private capital markets. We're posting on their behalf while they remain in stealth. Interested candidates should apply via LinkedIn — role details below. About us We are an early-stage AI company building decision intelligence for private capital. Our customers are private equity and venture firms across Europe, the UK, and the United States. Investment firms make high-value decisions over long horizons and learn remarkably little from the results. The context behind each decision is scattered across email, documents, and individual memory, and outcomes are almost never measured back against the original call. We are building the platform that changes that. We build across three sites — Dublin, Singapore, and New York — with Dublin as our AI focus and engineering based in both Singapore and New York. The role You will own the applied research that turns messy, real-world firm context into extraction, retrieval and reasoning systems investors will actually trust with a nine-figure decision. This is an applied role. The hard problems here are not novel architectures — they are grounded extraction from adversarially messy source material, entity resolution across a decade of unstructured records, retrieval under strict permission boundaries, and evaluation of systems where ground truth arrives years late. You will be as close to the evaluation harness as to the model. Build extraction and grounding pipelines. Lift terms, KPIs, covenants, financials and commitments out of IC memos, LP letters, Excel models, board decks, PDFs and email threads — every claim linked to its source document, version, author and access rights. No unsourced assertions, ever. Own retrieval quality. Design and tune retrieval over a heterogeneous, permissioned corpus spanning years of firm history. Get the right five documents in front of the model, not the plausible five hundred. Derive structure from unstructured work. Build the systems that infer a decision record — the reasoning, assumptions and confidence behind a call — from work already happening, so capture is a by-product rather than a form someone has to fill in. Build the evaluation infrastructure. Define what "correct" means for grounded extraction and decision derivation, build the golden datasets with domain experts, and make regression testing routine. Evals are a first-class deliverable, not an afterthought. Work on calibration and outcome scoring. Connect recorded decisions to realised outcomes; measure calibration at individual and firm level; surface where confidence and results systematically diverge. Stay model-independent. We are deliberately vendor-neutral across frontier and open-weight models. You will build so that better models make our layer more valuable, not obsolete. Sit close to users. You will watch real investment committees, valuation reviews and diligence processes, and take those observations directly into the system design. What we are looking for Required 8+ years building applied ML/AI systems in production, with meaningful recent work on LLM-based systems (retrieval, agents, structured extraction, fine-tuning or evaluation). Deep knowledge and understanding of large language models, neural networks and retrieval mechanisms — how they work, how they fail, how they are evaluated, and how they are actually applied in industry rather than in a paper. Deep, practical strength in Python and the modern applied AI stack. Heavy hands-on experience with LLM-assisted coding tools — you use them as a core part of how you build, and have views on where they help and where they don't. Demonstrated ability to take a genuinely messy real-world corpus to a system that people rely on — including the unglamorous work of data cleaning, entity resolution and error analysis. Rigour about evaluation. You can describe an eval harness you built and what it caught that vibes-based testing did not. Comfort with ambiguity and a strong bias to ship. This is a founding technical team with no established playbook. Clear written communication. You will write for engineers, for investors and for the record. Strongly preferred Experience in financial services, investment technology, legal tech, or another domain with high-stakes documents, strict permissioning and low tolerance for hallucination. Familiarity with knowledge graphs, entity resolution, or temporal/bitemporal data modelling. Published research, meaningful open-source contributions, or a PhD in ML, NLP, statistics or a related field. A PhD is desirable but not required — sufficient depth of hands-on experience is fine, and we care most about what you have shipped. Exposure to MCP, agent tooling, or building AI interfaces consumed by third-party systems. Experience with calibration, forecasting, causal inference or decision science. Not required Private equity experience. We will teach you the domain; several of the team came from outside finance. Curiosity about how investment decisions actually get made matters more than prior exposure. Why this role is unusual A research problem with a moat. Outcome-labelled institutional judgment accumulates only from the moment capture starts. The dataset you help build cannot be scraped or bought. Founding scope. Two of these roles exist. You will set technical direction, not inherit it. Frontier of the field. Connectors are becoming commodity. Making a firm's own context genuinely AI-usable, governed and traceable is not. Process Intro call with the hiring manager (30 min) Technical deep dive on your prior applied AI work (60 min) Practical exercise: grounded extraction and evaluation design, discussed live rather than take-home graded (90 min) System design and research judgement session (60 min) Founder conversation and team fit (45 min) We aim to close within three weeks of first contact. We are an equal opportunity employer. We are building a team from a deliberately wide range of backgrounds and will make reasonable accommodations throughout the process — just tell us what you need. This role is at a stealth-mode company. Further detail on the product and the backing is shared under NDA at the second stage. Interested parties, kindly apply through the link.