Principal Product ManagerData & ML Platforms
I build and scale machine learning products. 15+ years across e-commerce, adtech, and fintech — search, recommendations, customer data platforms. Most recently built Irrational Signals end-to-end — an ML trading-signals product — then made the call to retire it.
Numbers do the work.
- +20%search customer satisfaction
Multi-intent ranking framework. Takealot.
- 3×recommendation impression share
Flexible context-aware API. Takealot.
- Petabytecustomer data platform
Composable CDP, multi-brand, identity resolution. Takealot Group.
- 100+/hrtrading signals
Irrational Signals. US equities · 3 sectors · 6 batches per market day.
- 80% → 30%GeoIP-to-WiFi conversion target
Micro Networks model. Vicinity Media (proof-of-concept).
- 15+years
E-commerce, adtech, fintech. MBA. Cape Town.
A project execution framework for machine learning.
Built at Takealot Group to make the gap between exploration and productionisation explicit. The diamond below is the core visual: ideation expands into multiple candidate paths, then contracts into a single productionised model. Click any stage for detail.
The framework is about shape, not duration. Cadence varies with stack maturity, dataset readiness, and modern tooling (synthetic data generation, LLM-assisted prep, foundation-model baselines) often compresses several stages substantially.
Running the 6-month framework in practice surfaced two missing phases. The longer 15-month variant adds these as inline additions to the same diamond — not new structure, just acknowledgement that real ML work sometimes needs a dataset detour and a thinking pause.
- Learning 01Inserted before Generate Training Data
Dataset Exploration Phase
Often the bottleneck isn't the model — it's the dataset. Before committing to model training, generate multiple candidate datasets in parallel and pick the winner. Two evaluation paths: assess them via offline modelling against established metrics (when you'll need a model anyway), or — when the use case is simple enough — serve them directly as an A/B test. A caveat worth stating: if direct-serve datasets are sufficient on their own, you probably don't need a model at all.
- Learning 02Inserted after A/B Test, before retraining
Investigation Buffer
The A/B test doesn't always yield a clear answer. Reserve dedicated time after a test to investigate the result — understand the customer behaviour underneath the metric — instead of forcing immediate iteration. This stops cascading dependencies when an A/B test result requires deeper analysis (e.g. the Category Intent rollout, which needed investigation before promotion).
Both additions extend the project shape — a dataset arc up front, a thinking pause at the end. In practice they can add real time to a programme, though modern tooling (synthetic data generation, LLM-assisted analysis) compresses both phases substantially.
The diamond is calibrated for classical supervised machine learning — train a model against labelled data, evaluate offline, A/B test, productionise. The shape of the work still holds for agentic workflows: same expansion / contraction logic, same need for explicit checkpoints. Several stages just compress or change form.
- Generate training dataPrompt design · RAG corpus · synthetic data
- Candidate Model nCandidate prompts · tool sets · orchestration patterns
- Iterative trainingPrompt iteration · retrieval tuning · tool-spec refinement
- Offline POC evaluationLLM-as-judge · behavioural tests · red-team probes
- A/B test or internal testA/B with completion rate · human eval · latency budgets
- Productionise + stageGuardrails · rate limits · fallback prompts
- Drift monitoring+ hallucination rate · tool-call accuracy · latency
The two learnings (the dataset detour and the investigation buffer) carry over directly. The investigation buffer matters more in agentic work, not less — behavioural failures rarely have a clean numeric answer.
These are the artefacts that typically accompany a project on this framework. Names and roles only — the actual templates and specifications aren't included on this site; they live with the team that adopts the framework.
Career
- 2025 — archived
Founder & Principal Product Manager
Irrational SignalsBuilt and ran an API-first, ML-driven trading-signals product for US equities end-to-end — data pipeline, LightGBM model, REST API, Python SDK, and billing. Validated live, then deliberately wound down.
FintechMLAPI products - 2024 — 2025
Group Product Lead — Customer Data & ML Platforms
Takealot GroupEnd-to-end product strategy for the Customer Data Platform, CRM, fraud, and insights across the multi-brand portfolio. Architected petabyte-scale data platform and scalable analytics architecture for cross-brand activation.
Customer Data PlatformIdentity resolutionMulti-brand - 2022 — 2024
Group Product Lead — AI/ML (Search & Recommendations)
Takealot GroupMachine-learning product strategy for search and recommendations across platforms. Multi-intent search framework (+20% customer satisfaction). Flexible recommendations API (3× impression share).
SearchRecommendationsML platforms - 2020 — 2022
Product Manager — AI/ML (Search & Recommendations)
Takealot.comDelivered the machine-learning product roadmap for core search and recommendation performance. Defined and tracked North Star metrics. Productionised models with data scientists and ML engineers.
SearchRecommendations - 2017 — 2020
Product Lead — AdTech Platform
Vicinity MediaOwned the advertising-tech platform strategy: audience data, campaign systems, attribution. Built the data science function. Out-of-home (OOH) attribution linking offline ad exposure to in-store behaviour. WiFi-clustering and point-of-interest visit models.
AdtechAttributionAudience - 2008 — 2015
Consulting Manager
Cape Value GroupData analysis and market modeling across financial services and public sector. Business strategy, forecasting, and operational planning.
ConsultingModeling
Selected work, in depth.
- Case 012020 — 2024
A multi-intent modelling framework for search
Reframing ranking from a single black-box model into composable intent models — the FY23–FY25 vision.
SearchML Platforms3-Year VisionTakealot GroupRead → - Case 022022 — 2024
Flexible recommendations: a context-aware API
One API serving every surface — PDP, cart, wishlist, search, home. Context-aware delivery from a composable model library.
RecommendationsPlatform ArchitectureAPI DesignTakealot GroupRead → - Case 032024 — 2025
Composable customer data platform
Petabyte-scale, first-party tracking events, cross-device and cross-brand identity resolution. Multi-brand activation.
Customer Data PlatformData PlatformIdentity ResolutionTakealot GroupRead → - Case 042017 — 2020
Micro Networks: a proof-of-concept for clawing back precise location
A POC model designed to extend Vicinity's existing location accuracy into the 80% of ad requests where users didn't share their coordinates.
AdTechLocation accuracyProof of conceptVicinity MediaRead → - Case 052025 — archived
Irrational Signals: an ML product, end-to-end — and the call to retire it
An intraday trading-signals product for US equities, built and operated solo across data, model, API, and go-to-market. It ran live, then was deliberately wound down. Archived build case study.
FintechML / LightGBMAPI-firstFounderRead →