Emil Chetty
Conference affiliation: Wise
Emil Chetty leads product strategy for data and AI platforms at Wise, where he was introduced as Group Head of Data & AI at AI Engineer NYC 2026. His work spans foundational data infrastructure, analytics, machine learning and generative AI. His approach to payment agents centers on a specific division of responsibility: customers approve the recipients and financial terms, agents choose how to carry out the work, and payment services enforce what customers approved.
On this page
Emil has more than a decade of experience building production data and AI systems across financial services, public transport and open knowledge. His platform work includes streaming data, machine learning infrastructure, LLM gateways and agent runtimes. At Wise, he works with engineering, security, privacy, risk and compliance teams on evaluating, deploying and operating those systems.
Shared infrastructure for Wikimedia’s data
Before Wise, Emil was a lead technical product manager at the Wikimedia Foundation and led the Metrics Platform product effort. The team developed shared services and client libraries for collecting and using product data. Common handling of session identifiers, timestamps and sampling let web and mobile teams instrument their applications without solving those problems separately.
His work also addressed measurement for organizations contributing material to Wikimedia Commons. In 2022, Wikimedia Israel’s analytics dashboard faced growing storage costs that constrained its ability to serve additional cultural institutions. Emil helped work through the options with its technical lead. Fetching view counts through an API could reduce duplicated storage and processing, but the required refactoring exceeded the maintainer’s near-term capacity. He coordinated exploration of Wikimedia cloud hosting and pressed for broader prioritization of cultural partners’ measurement needs. Wikimedia Israel subsequently completed the migration; Emil’s contribution was product coordination and technical scoping within that collaborative effort.
AI applications for employees
At Wise, Emil worked alongside ML platform lead Mark Harley on introducing Credal for internal AI applications. The team wanted access to multiple model providers, APIs that could connect to audit logging and cost tracking, and an interface employees could use without building an application themselves.
That combination supported a progression from asking questions of internal documents to creating reusable assistants and sharing them through Slack. Wise’s team applied the platform to customer-support quality reviews and to converting investigators’ internal accounts into structured suspicious-activity reports with human oversight. These applications give concrete context to Emil’s platform responsibilities: supporting employee-created tools while retaining the operational controls needed for support and compliance work.
Customer approval that payment services can enforce
In Stop Making Your Customers Click Approve, Emil argues that repeated permission requests can turn customers into coordinators of the work they intended to delegate. A payment agent needs more than a judgment that an action looks safe: it needs authorization from this customer to pay this recipient on these terms. He retains a role for evaluations and agent harnesses, while separating them from the customer’s agreement.
His design makes that distinction concrete:
- Approve a defined job. In a hypothetical $775 payment run for three suppliers, preparation begins with read-only access. Trusted software resolves recipients, bank details and amounts from account records, and missing or conflicting information must be settled before approval. The customer sees the payment terms and the decisions being delegated. Authentication and confirmation bind that approval to a specific, immutable version of the job. Execution permissions are limited to the overlap between the approved job, Wise’s policies and the customer’s existing rights.
- Check payments outside the model. Payment services compare each action with the approved record. If a malicious instruction persuades an agent to use a different payee, the service rejects the changed recipient. Enforcement therefore does not depend on the model recognizing the scam, and the agent cannot rewrite its own authorization.
- Distinguish a retry from changed terms. A restarted worker can inspect the existing payment state and resume the same job without creating a second payment. An expired exchange-rate quote requires a different response: the agent prepares a revised proposal, shows the customer what changed and requests approval. The plan can change within the agreement; new financial terms require a new customer decision.
1 conference talk
Loading more talks and highlights…
