Roles & Responsibilities Own the architecture and technical roadmap for the Operational Intelligence Gen AI platform, defining reference architectures and standards that teams across the program can reuse and scale. Architect the enterprise tool-calling layer, setting standards for how internal services, data platforms, and external APIs are exposed to agents as tools. This covers schemas, contracts, versioning, access control, and resilient execution (parallel and chained calls, retries, timeouts, idempotency, circuit breakers, and fallbacks). Set the architecture for multi-agent and skills-based agent systems (e.g., LangGraph, CrewAI, AutoGen), including when to use agents versus deterministic workflows and how agents coordinate, delegate, and share context. Design self-correcting agent patterns: tool critics and verifiers that check tool selection, arguments, and results, plus reflection and plan-and-execute loops that catch errors before agents act or respond. Architect agent memory and state for long-running, multi-step sessions, including short- and long-term memory (e.g., Mem0, Zep, LangGraph memory), session persistence, and memory scoping and retention. Define retrieval architecture for grounded reasoning, including Self-RAG, Corrective RAG (CRAG), adaptive retrieval, and Graph-RAG, built on vector databases (e.g., Pinecone, pgvector, OpenSearch) and knowledge graphs (e.g., Neo4j, AWS Neptune). Design the inference and orchestration layer for high throughput, covering async processing, streaming, caching, rate limiting, and model routing across OpenAI, Anthropic, Gemini, and open-source models. Establish the LLMOps architecture on Databricks (AWS), including MLflow, CI/CD for prompts and agents, and model lineage. Set platform-wide evaluation and observability standards: end-to-end tracing of tool calls, tool-selection accuracy, task success rates, cost, and latency. Build security and Responsible AI into the architecture, including guardrails, prompt-injection defenses, hallucination mitigation, data privacy, and audit trails that meet Mastercard's data ethics and regulatory standards. Own key architecture decisions and trade-offs, such as build versus buy, model and vendor selection, and cost versus latency versus accuracy, documented through architecture decision records. Assess emerging models, frameworks, and standards such as MCP for production readiness. Lead design reviews and partner with engineering, product, security, and data teams to align on architecture and raise the engineering bar across the program. Stay hands-on, building prototypes and reference implementations that de-risk decisions and give teams a working starting point. All About You Bachelor's or Master's degree in Computer Science, AI/ML, or Engineering, or equivalent practical experience. Significant experience architecting large-scale distributed systems in production, including microservices, event-driven systems, data platforms, and cloud-native infrastructure. Deep expertise in API design and systems integration, including contracts, security models, and reliability patterns for services consumed by many teams. You understand what changes when the consumer is an LLM. Hands-on experience designing and shipping LLM-powered systems in production, especially agents that use tool/function calling, memory, and retrieval, and a clear view of how they fail. Strong Python skills for hands-on prototyping (async programming, FastAPI or similar), plus SQL proficiency. Deep experience with at least one major cloud platform (AWS preferred; Azure or GCP welcome), including its AI services such as Bedrock, SageMaker, Azure OpenAI, or Vertex AI. A solid grounding in security, privacy, and compliance for enterprise systems, ideally in financial services or another regulated industry. The ability to influence without authority, aligning engineering, product, and leadership around a technical direction and explaining trade-offs in business terms. Nice to have: hands-on experience with tool critics, Self-RAG, CRAG, reflection loops, memory frameworks such as Mem0 or Zep, MCP tool servers, or agent evaluation frameworks. Nice to have: exposure to fine-tuning (LoRA, PEFT, RLHF), multimodal models, or knowledge graphs. Coming from backend or distributed systems architecture? The hardest problems in enterprise Gen AI are reliability, integration, security, and scale, and you've solved them before. If you've started applying that experience to LLM systems, we want to hear from you, even if you haven't used every tool listed here.
Corporate Security Responsibility
All activities involving access to Mastercard assets, information, and networks comes with an inherent risk to the organization and, therefore, it is expected that every person working for, or on behalf of, Mastercard is responsible for information security and must:
Abide by Mastercard’s security policies and practices;
Ensure the confidentiality and integrity of the information being accessed;
Report any suspected information security violation or breach, and
Complete all periodic mandatory security trainings in accordance with Mastercard’s guidelines.