AI Pulse: From FX forecasting to cheaper, more personal AI tools

Ant International’s upgraded FX model, FalconTST 2.0, has attracted Citi, HSBC, and Standard Chartered. The focus is practical: helping businesses forecast currency movements, manage exposure, and plan cash flow. For builders, this is a reminder that specialised models tied to a clear business workflow may matter more than general-purpose demos.

Starling Bank’s new Smart Tools let customers automate tasks such as budgeting and tax preparation, while also creating their own AI-powered banking features. That raises an important design challenge: useful personalisation needs to be paired with clear permissions, transparency, and safeguards around sensitive financial data.

Nvidia’s open-source NeMo Switchyard is designed to route requests to less expensive models, with the report describing a potential 74% cost reduction in exchange for a 6% accuracy trade-off. For teams running AI at scale, intelligent routing could be as important as choosing the model itself—but only if users can decide where lower accuracy is acceptable.

Google’s latest Pixel lineup includes slimmer camera hardware, stronger zoom, and additional AI features. The broader shift is familiar but significant: AI is increasingly being presented as part of the phone’s everyday camera and software experience, not as a separate app users must seek out.

TrueFoundry has introduced TrueForge, a vendor-neutral agent harness paired with its AI Gateway. It is positioned for teams that want to build and govern production agents across different models or MCP servers, with a reported 50% lower cost than Claude managed agents. The appeal for engineering teams is flexibility; the open question is how much operational work that flexibility creates.