This is not a traditional CFO practice with a Python side project. Every analytical capability listed here is grounded in 61 working models from the MIT Applied AI & Data Science Program—and deployed directly in client engagements alongside board-level finance leadership.
Each capability below maps a specific machine learning technique to a finance outcome—with the methodology, the tools, and the business context all understood by the same person deploying it.
Classification models trained on historical customer data to predict future behaviour. In a SaaS context this means a weekly churn risk score per account—a probability that a given customer will cancel in the next 30, 60, or 90 days. Features include login frequency, feature adoption rate, support ticket volume, payment behaviour, and contract age. XGBoost and Random Forest consistently outperform simpler models on real SaaS datasets.
The same framework applies to attrition risk for PE portfolio companies, lead conversion probability from CRM data, and loan default risk in financial services. The model output is always actionable: a ranked list of accounts or individuals requiring intervention, with the probability score quantifying urgency.
Unsupervised learning that finds natural structure in customer data without telling the algorithm what to look for. K-Means and hierarchical clustering reveal segments based on behaviour, contract value, product usage mix, or acquisition channel. The resulting segments drive pricing strategy (which tier to price up), upsell targeting (which segment is ready for expansion), and renewal prioritisation (which segment is habitually at risk).
PCA and t-SNE reduce high-dimensional customer feature sets to two or three dimensions for visualisation—making segments legible in board presentations rather than buried in a spreadsheet. Applied to AllLife Bank customer segmentation and marketing campaign analysis in MIT capstone work.
Large language models applied to finance document workflows. The most mature use case is automated management commentary: actuals data is extracted from the ERP, variance calculations are computed in Python, and an enterprise LLM API generates the narrative—month-end commentary, variance explanations, board pack summaries—ready for CFO review. The finance team edits rather than writes from scratch.
Beyond reporting, LLM pipelines handle document summarisation (FDD reports, board papers, contracts), financial query answering on structured data, and quality assessment using an LLM-as-judge framework. Prompt engineering patterns—zero-shot, few-shot, chain-of-thought—are tuned to the finance domain for reliable, consistent output. Demonstrated in the MIT capstone Real-Time Retail Feedback Intelligence project.
An 8-course, hands-on program taught by MIT faculty. Not an academic credential held at arm’s length—61 working Jupyter notebooks covering the full ML stack from statistical foundations through deep learning and generative AI, all applied to finance-equivalent problems.
LLM pipeline using Llama-2 for banking customer intent classification. Foundation for automated board commentary and document summarisation workflows.
NumPy, Pandas, probability, hypothesis testing, and statistical inference. Underpins all analytical work—financial data is just data, and these tools apply directly.
PCA, t-SNE, K-Means, DBSCAN, GMM, hierarchical clustering, and network analysis (NetworkX). Applied to customer segmentation and relationship mapping.
KNN, logistic regression, SVM, decision trees, cross-validation. Employee attrition prediction—same classification framework as SaaS churn modelling.
Random Forest, XGBoost, AdaBoost, Gradient Boosting—applied to classification problems including churn prediction, loan default, and employee attrition. The ensemble framework underpinning the SaaS churn risk scoring model.
ANNs, CNNs, Transformers in PyTorch. BERT fine-tuning for text classification. Transfer learning (VGG16, EfficientNet). Applied to financial document classification and text analytics.
Collaborative filtering (KNN, SVD), matrix factorisation. Applied to SaaS product affinity and cross-sell targeting.
Loan default prediction (own work, Random Forest). Marketing campaign segmentation (K-Means/GMM). Retail Feedback Intelligence—LLM pipeline with zero-shot, few-shot, and chain-of-thought prompting.
Each example below is a concrete deployment—not a proof of concept, but a working system built alongside CFO-level finance leadership in a real business.
A Random Forest or XGBoost classifier is trained on 12–24 months of customer behaviour data: login frequency, feature adoption, support ticket volume, payment behaviour, and contract age. The model is re-scored weekly and produces a churn probability per account.
The output is a ranked list of at-risk customers—not a chart, but a prioritised intervention queue for the customer success team. The customer who was 8% risk last week and is 34% this week gets a call before the one who has been flat at 15% for a month.
Actuals data is extracted from the ERP via API or CSV export. Python calculates the variances—actual vs. budget, actual vs. prior year, current month vs. prior month. These variances are passed to an enterprise LLM API with a finance-domain system prompt and few-shot examples of acceptable commentary style.
The API returns a first draft of the management commentary section—department by department, variance by variance. The CFO edits rather than writes from scratch, saving 2–4 hours per month-end cycle and improving consistency across reporting periods.
Most finance functions know they should be doing more with AI. Far fewer get there without wasting the budget. The projects that fail tend to fail the same way—handed to technologists who have never run a month-end close, or to finance teams who do not know what a model can and cannot do. That gap is where the money disappears.
Graeme sits on both sides of that gap: 20 years running finance functions, and MIT-trained, hands-on machine learning capability. That means an AI transformation of your finance function can be scoped, built, and embedded by one person who understands the accounting, the data, and the models—and who will tell you honestly which problems AI actually solves and which are better fixed with a good spreadsheet.
Graeme scopes the opportunities, builds the models and automations himself, and embeds them into your monthly cycle—end to end. Suited to businesses that want the outcome without first hiring and managing a data science team to get it. You are left with working tools, documented so your finance team can run them long after the engagement ends.
Already have a data hire, an analytics team, or an external vendor? Graeme provides the finance-literate oversight that keeps the programme honest—prioritising what is worth building, reviewing the technical approach, and translating between the board and the builders. The person who can challenge a data scientist on the model and a CFO on the use case in the same meeting.
Every transformation starts with a short diagnostic—where AI genuinely adds value in your finance function, prioritised by return, and delivered in increments rather than as a big-bang project.
Croftlands does not just advise on AI adoption—it runs on it. Every analytical tool, financial model, and client deliverable is built and maintained using modern AI coding assistants and enterprise LLM APIs as active collaborators in the work.
Most CFOs who talk about AI are referencing a vendor product they have seen demonstrated. This is different. Enterprise LLM APIs are used to build the finance tools themselves—generating Python code for churn models, automating data pipelines, and producing first drafts of board commentary. An AI coding assistant is used to build and maintain the analytical infrastructure.
This means two things for clients. First, the turnaround on analytical tools is faster than traditional consulting—a churn scoring model or cashflow forecast can be prototyped, tested, and deployed in days rather than weeks. Second, the tools are explainable: because the CFO understands the code and the methodology, they can explain to your board and your PE investor exactly how the churn probability was calculated and what assumptions drive the forecast.
It also means Croftlands can advise on your own AI adoption strategy from a position of genuine practical experience—not from a sales deck, but from running these systems in a live finance practice.
Start with a conversation about what your data looks like and what questions you need answered. The right model follows from there.
Book a Discovery CallOr reach out directly: graemeweeden@croftlandsconsulting.com