David Abitbol describes Revolut’s evaluationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. of AI coding tools using measures of coding time and lead time across a sample of engineers. He says the company moved toward self-hosted open models to reduce costs and provide broader access.
He discusses GPU capacity, time to first token and tuning the same hardware to serve more requestsAI inference throughput measures how much model-serving work a system completes per unit of time.. The talk covers trade-offs involving context length, quantisationModel quantization represents AI model values with fewer bits to reduce storage, memory use and often inference cost. and caching, and describes routing image tasks to a separate modelAI model routing sends each request to a model chosen for that request's complexity, cost, speed, privacy, or other requirements..
Abitbol outlines how the company approaches operational controls, including sandbox rulesAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access., human approval and observability. His performance and cost figures are presented as Revolut’s reported experience rather than general guarantees.
Watch on YouTube
