Founder and lead engineer

AI and data solutions for financial services firms, Cape Town

• Rebalancing engine, REIT fund manager. Multi-currency NAV and weights for each book, in a pipeline now running with the portfolio team alongside their manual process to validate outputs. Turning the manager’s morning routine (blend the books’ drift, round each line to the nearest 5 bps, let cash reconcile to 100%) into suggestions approved line by line. Scoped LSEG API ingestion for end-of-day pricing.
• KYC automation, asset manager. A tool that automates client onboarding and KYC document checks.
• Debt-review correspondence, prototype. For a firm whose creditor replies were keyed into the CRM by hand: reads each email, matches it to the right consumer account, extracts balances with the source sentence attached and holds them for a person to approve. When it isn’t sure, it names its candidates instead of guessing.
• Research platform, asset manager. Built on a fork of the open-source TradingAgents framework and extended into a research platform whose answers cite their sources: a resolver that checks client documents first, then licensed data, then public sources, retrieval over vector embeddings and a job worker for long runs. Code on GitHub
• Data reliability platform. Where Pallisade began: an AI copilot for data pipelines that answers questions about the warehouse, traces lineage across the stack and catches broken data before dashboards do. Built on Airflow and dbt. Code on GitHub
• Operations system. The internal system I run Pallisade on: leads, outreach, follow-ups, projects and AI summaries. Code on GitHub
• Peer review platform, engineering group. Building a web platform to manage and track peer reviews.
• Delivery and governance. Map client processes before building, scope AI features with non-technical stakeholders, then build, deploy and hand over. Company POPIA Information Officer. Run public “AI in Finance” webinars.

Co-founder and technical lead

Marketing automation agency for private medical clinics

• Clinic automations. Workflow automations on GoHighLevel for lead capture, bookings and follow-ups, plus CRM automation and email deliverability.
• Conversion tracking. Server-side Meta Conversions API tracking, so clinics could tie bookings back to the campaigns that drove them.
• Client products. A dental practice platform, a veterinary communications tool and a fitness coaching app.

PGDip research project

A controlled experiment on LLM advice for broken dbt pipelines, building on Pallisade’s data reliability work. Fifteen failures injected into a dbt and DuckDB project, prompts at three levels of context, three models (two hosted, one local Llama 3) and three repetitions: 405 calls. Answers are scored blind and by two experts, with agreement measured by Cohen’s kappa, and checked automatically for hallucinated references. Analysis: two-way ANOVA per model with Tukey post-hoc tests and Holm correction, with Friedman tests as a rank-based check. First result: more context makes all three models’ fixes more specific.

PGDip coursework, NumPy and Spark

Predicting Amazon’s Best Seller badge across 133,102 Kindle books, where only 1.7% carry it. Logistic regression written in NumPy, with gradient descent and Newton-Raphson and verified with five checks. Features chosen with variance inflation factors and backward elimination on Wald p-values; models judged on PR-AUC, because accuracy rewards flagging nothing. Class weights were tested rather than assumed, and rejected because they hurt the ranking. With a threshold tuned on validation data, one in nine flagged books is a bestseller, against one in 59 by chance.

PGDip coursework, TensorFlow and Spark

Naming the author of a 180-word passage from style alone: 21,000 passages from 60 novels by six authors, split by book so no novel appears in both training and test. A TF-IDF baseline, then three LSTM versions, each change tied to a diagnosed weakness. Removing character names took away an easy shortcut, and a Bi-LSTM 74% smaller than the one before scored higher: 0.685 macro-F1, against 16.7% accuracy from always guessing one author.

Deep-learning medical-imaging competition, Efficiency Track

Balancing model accuracy against latency and compute cost.

Payments product design

A fuel-price-lock product: users prepay for litres at today’s price and redeem them with a card restricted to fuel merchant category codes. Issued under a sponsor-bank programme-manager model, with price risk hedged through an OTC swap. I mapped its regulatory requirements across SARB, PASA, FICA, FAIS and POPIA, and built the app prototype. Code on GitHub