Capital One isn't just adopting AI; they're actively building it. Their strategy for a robust, scalable multi-agent AI platform centres on customising open-weight models with their unique proprietary data, rather than solely relying on off-the-shelf foundational models. This approach is built upon years of investment in data transformation and cloud adoption, providing a solid technical foundation that allowed them to move swiftly when the latest AI wave hit.
The core of their innovation lies in deeply fine-tuning open-weight models. Capital One views its extensive, proprietary data as a significant competitive advantage that general frontier models can't replicate. By feeding this data into op en models, they are effectively training AI agents that are experts in Capital One's specific use cases, policies, and terminology. A surprising benefit of this customisation is its enterprise-wide extensibility; training a model for one task often yields a general performance lift across their entire portfolio.
This multi-agent architecture, dubbed MACAW, has been instrumental in handling complex customer interactions. For instance, a fraud detection workflow manages millions of calls annually, where a single large language model proved insufficient. MACAW breaks down these interactions into specialised agents: an understanding agent to decipher customer intent, a reasoning agent to summarise, a validation agent to fact-check, and an explaining agent to format the output. This not only streamlines processes for human agents but also automates tedious post-call documentation. Capital One's Chat Concierge, an auto-shopping assistant, also employs a similar agentic division of l abour, leveraging a customised version of Meta's Llama model.
Beyond customer-facing applications, Capital One is using agentic AI to automate internal tasks, such as optimising backend hosting infrastructure. An autonomous system searches for the best configurations to achieve optimal latency, freeing up researchers for higher-level work. Looking ahead, Vanee anticipates a rise in model routing abstraction layers for enhanced accuracy and cost-effectiveness, as well as a significant shift towards proactive, event-driven AI that acts without explicit human prompts, requiring rigorous testing and monitoring.
Fuente Original: https://venturebeat.com/orchestration/why-capital-one-built-its-multi-agent-ai-platform-around-open-weight-models
Artículos relacionados de LaRebelión:
- Nvidia Embraces Open-Source AI to Boost Spending
- Nvidia Launches Its First Open-Source AI Model
- AI Agents Coordinating in Real-Time Outperform Advanced Models
- Texas Halts Data Center Grid Links Amidst Demand Surge
- Claude Mythos 5 Puerta trasera en proyecto open source
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario