My career in machine learning started in research—filing patents and publishing papers at Adobe Research Labs while completing my graduate studies. That experience shaped a core conviction I still hold today: research that doesn't reach production is unfinished work.

Over the past 15+ years, I’ve built and scaled artificial intelligence (AI) systems across high-stakes domains. As Chief Technology Officer (CTO) at GIST Impact, I built machine-learning (ML)-powered platforms for regulatory risk intelligence, processing compliance signals across more than 20,000 companies. At Meta, I drove ML and product strategy across the multi-billion-dollar ads ecosystem, focusing on personalisation and ranking at a global scale. Along the way, I’ve filed 8 US patents, published 9 papers, and authored a book on the mathematics of AI and ML.

When evaluating my next move, Wise stood out for one clear reason: the scale and depth of the ML challenge. Wise moves billions across borders for millions of customers and the financial crime landscape is evolving fast - deepfakes, synthetic identities, coordinated networks. Throughout my conversations with engineers, data scientists, product managers and our CTO, Harsh, one thing became immediately obvious - a shared conviction that staying ahead of these threats is fundamentally a machine learning challenge, paired with the ambition to build a world-class platform to do it.

The product complexity at Wise is real - we operate across teams in Fraud, Anti-Money Laundering, Sanctions, KYC & Compliance, each with distinct regulatory frameworks, detection logic and operational workflows. As Wise expands into new markets and  products, the platform needs to support a wider range of risk and specialised models. That creates an interesting engineering challenge around making those models easier to develop, evaluate and deploy. We’re building shared infrastructure for training and deployment, and exploring approaches including deep learning and graph-based models.  

That's not a linear scaling problem. It’s about building a platform that helps teams turn new ideas into tested detection capabilities.

What are the hardest technical challenges your team is tackling that feels like a major win for both the customer and the mission?

There are a few that I find genuinely fascinating and they're the kind of problems that don't exist at most companies:

Graph intelligence: Graph connections are the strongest signals in financial crime - risk propagates through networks in ways that transaction-level models can't see. We're building unified Graph Intelligence: one authoritative graph computing PageRank, community detection, label propagation and composite edge scoring - serving graph-native features into our Feature Store for every downstream model. Along with Graph Neural Networks (GCN, temporal, GraphSAGE) for learned representations that capture structural patterns invisible to tabular models.

Detection architecture for 10x scale: As Wise grows, our detection models need to scale across new crime typologies, new architectures and billions of training rows across hundreds of countries. We're investing in deep learning, sequence models, distributed training and a model platform where deploying a new specialised model is a config-driven pipeline. The architecture choices when you're training at that scale get genuinely interesting.

Ground truth at scale: In FinCrime, labelled data is expensive, noisy and carries a direct regulatory surface. We're evolving a Label Platform - a versioned, auditable ground-truth system with label quality monitoring, confidence scoring, programmatic labelling and label propagation that can surface previously unseen risk patterns from known signals. Investigation outcomes feed into our Label Platform as governed ground truth - with quality monitoring, confidence scoring and version control - continuously improving the models that power our detection systems.

We're also applying LLMs to red-flag detection, investigation summarisation and mining historical case narratives to generate structured training data - with real engineering challenges around evaluation frameworks, hallucination detection and confidence calibration.

What does a typical day look like for someone on this team?

Honestly, it depends on which part of the platform you're working on. If you're on the Feature Platform, you might be designing a new graph-native feature - for instance, computing multi-hop risk density from a graph DB, benchmarking it against the existing feature set and shipping it to the Feature Store where downstream models automatically pick it up.

If you're on Risk Learning, you could be building the label quality monitoring pipeline - detecting distribution drift in training labels, flagging when a typology's precision degrades, or designing the active learning loop that identifies which cases to route for human review.

If you're on the Modelling and Risk ML platform side, you might be building declarative training pipelines that scale across billions of rows, designing the evaluation framework for a new typology-specific model, running GNN experiments on our graph data, or working with the MLOps pipeline to get a model from experiment to shadow-mode scoring fast.

Across all of it, there's deep cross-functional collaboration - with domain experts in Fraud, AML, Sanctions and Compliance teams who bring the investigative intuition, with Data Scientists who bring statistical rigour and with infrastructure engineers who keep it all running at scale. The problems are genuinely hard, the data is rich and the feedback loop to real-world impact is short: when a model catches a mule network or a new fraud typology, you know it.

This is a team where you own the full lifecycle. You design the system, build it, ship it to production and see it catch bad actors out in the real world. We operate with empowered teams that own their problem space end-to-end.

To give you a concrete sense: we're building the next generation of our Graph Intelligence Platform, Label Platform and Model Platform. These are foundational systems that will underpin every FinCrime detection model at Wise. If you join now, you're shaping the architecture at a defining moment - the graph feature store schema, the label versioning model, the declarative training pipeline, the GNN architecture choices - these are decisions being made now, by the engineers on the team.

We operate an experiment-driven culture - ideas get tested with real data, validated with real metrics and shipped when the evidence supports it. You'll have the rare combination of a hard unsolved problem, the data to attack it and the ownership to define how.

If you're the kind of engineer who gets excited by technically hard problems with direct human impact, where your model doesn't just optimise a metric but actually stops someone's savings from being stolen, this is the place to be.

To thrive here, you need to combine deep engineering rigour with a sense of who you're building for. Behind every transaction is a real person trusting Wise with their finances. You also need comfort with ambiguity - FinCrime detection is adversarial and the right architecture for a problem isn't always obvious. We’re at our best when we solve problems together. That means bringing real evidence to the table, being honest when we're unsure, and following through until our ideas actually make a difference.

If that sounds like how you want to work, we're hiring!