I validate and build AI and quantitative models for financial services.
My day job is independent model validation and AI governance: testing whether models, including LLM-based ones, do what they claim, and whether a regulator would agree. On the side I build AI tools, and I try to hold them to the same standard: use the model for judgement, then make the code check it.
Featured work
- wealth-advisory-validation - How I test the models behind robo-advice: risk profiling, suitability and portfolio risk. 50 tests, each tied to a regulatory requirement and the NIST AI RMF, with 26 worked examples on simulated data.
- llm-agent-validation - The same approach applied to LLM agents in financial services: sycophancy, red-teaming with confidence intervals on failure rates, checking the LLM judge itself, and calibration. 25 tests and 14 worked examples.
- nlp-paper-critiques - Six critiques of published NLP research, written for Cornell CS 4740 (2021). They cover deception detection, sarcasm, bias in word embeddings, style transfer and event extraction.
- ticket-scenario-matching - An NLP project from 2023–24: matching service-desk tickets to known problem scenarios without labelled data, using BERT to mask sensitive text, four embedding models and a separate threshold for each scenario. It includes a review of the project as I would validate a client's model, and an erratum: a merge step had silently dropped 15 scenarios.
- nlp-in-the-financial-market - My part of a Cornell MPS team project (2022) with Gravity Investments on news sentiment and stock prices. I proposed and built the time-series analysis: Granger causality and cointegration tests, then VAR and VARMA models compared stock by stock.
- stability-selection-in-logistic-regression - My BSc thesis (TU Dublin, 2020): a Monte Carlo comparison of the likelihood ratio test, stepwise selection, the lasso and stability selection in logistic regression. It includes an erratum for a loop bug I found later, which explains the thesis's one unexpected result.
- lse-data-analytics-career-accelerator - Three LSE projects in SQL and Tableau, Python and R, covering marketing, NHS appointments and retail reviews, with the original reports, code and data. The NHS project includes an erratum: same-day appointments are close to half the total, where my report said a fifth.
- lp-financial-model-net-interest-income - A linear programming model of a company's month-by-month borrowing and lending, from my NIT Calicut final-year project (2013) and a 2017 paper with Prof. T. Radha Ramanan (IIM Indore). It includes the original Excel model and an erratum: one mis-referenced cell explains the gap between the paper's validation result and the company's own figure.
Earlier work: Finnish elections as a predator-prey system · Statistical network analysis with ERGMs
Toolkit: Python · SQL · R · LLM APIs · scikit-learn · pandas · Tableau · model validation (SR 26-2, formerly SR 11-7; PRA SS1/23) · AI governance (EU AI Act, NIST AI RMF, NIST AI 600-1 Generative AI Profile)
Background: FRM · MPS Applied Statistics, Cornell · BSc Mathematics, TU Dublin · LSE Data Analytics Career Accelerator