In production
MCP and agentic framework for the finance platforms, end to end
Defined and built the path an agent call takes into PayPal's finance systems — SAP platforms, Google Cloud, Kubernetes, and Linux infrastructure — from auth to the app: MCP server, federated identity assertion, policy guardrails, then the enterprise system, with a full audit trail persisted at every step and the resulting UX and UI shipped in the app. It is now the standard pattern teams follow for agentic workloads.
- Adopted as the standard pattern for connecting LLM agents to backend systems, and the reference architecture I use when enabling new teams.
- Proven first across SAP HANA, OData, and SaaS landscapes, then extended to the Google Cloud, Kubernetes, and Linux estate, cutting new integration build time from weeks to days.
- Identity travels with the request end to end, so access is attributed to the real user rather than a shared service account.
- Owned through to the product surface: the agent UX and UI in the app, not just the server behind it.
MCP · Federated identity (SAML, OAuth) · Policy guardrails · SAP HANA · OData · Google Cloud · Kubernetes · Linux · Agent UX / UI
In production
MCP framework, shared SDKs, and team enablement
An end-to-end framework with SDK libraries that absorb federated auth, query governance, and security guardrails, so teams building their own MCP servers inherit production-grade safety by default instead of re-implementing it per system.
- First to run the FastMCP runtime in production; the framework now underpins the MCP servers built on it.
- CLI-generated, production-ready MCP servers plus an interactive documentation site turned one-off builds into a repeatable authoring path that other teams follow without hand-holding.
- Governance is a property of the framework, not a review step, which is what makes the integration pattern above reusable.
FastMCP · Python · SAML / OAuth · Query governance · Guardrails · CLI tooling
In production
Enterprise GenAI adoption for everyday work
Driving the adoption of generative AI in the day-to-day work of employees and managers, not just engineering teams. The experience is built from curated MCP servers, plugins, and skills that connect the assistants people already use to the systems their work actually runs on, through the same governed path as every other integration.
- Curated MCP servers that expose the handful of tools a role needs, with descriptions and guardrails tuned so a non-engineer gets a reliable result without prompt craft.
- Plugins and skills that package common workflows for employees and managers as repeatable, governed experiences rather than one-off chats.
- Adoption treated as an architecture problem: identity, data governance, and audit are inherited from the framework, so rollout to a new group is a curation exercise rather than a security review.
Curated MCP servers · Plugins · Skills · Enterprise assistants · Governed tool surfaces · Change enablement
In use
Evals: the gate between a working demo and a shipped agent
Evaluation harnesses that decide whether an MCP server or agent is ready to ship: whether the model picks the right tool, whether the task actually completes, and whether a change regresses either. Run before release and on every change, so "it worked in the demo" is never the acceptance criterion.
- Tool-selection and task-completion evals for MCP servers built on the framework, with deterministic scoring so results are comparable across runs and models.
- Regression gates for tool descriptions and schemas, where most agent failures originate, so a wording change cannot silently break routing.
- The same discipline applied in public: the mlx-lab on-call eval below compares a local model with a frontier model on identical tasks, with every transcript reproducible.
Agent evals · Tool-use scoring · Regression gates · Deterministic harnesses · MLX
In progress
Knowledge graphs as the context layer for enterprise agents
Architecting the knowledge-graph layer that gives agents grounded context about enterprise systems, their relationships, and their owners, so an agent reasons over verified structure instead of guessing from retrieved text. Exposed to agents through governed MCP tools, with the same identity and audit path as every other backend.
- Graph-backed context for agentic operations: which system depends on which, who owns it, and what changed, queryable by an agent through a curated tool surface.
- Designed as a complement to retrieval, not a replacement: the graph supplies structure and relationships, retrieval supplies documents, and the agent gets both through one governed path.
- Informed by a close read of how vendors are positioning graphs and context engines, written up in the SAP NorthStar analysis under Writing.
Knowledge graphs · Context engineering · MCP · RAG · Enterprise metadata
In productionFleet in rollout
AIOps: agents for platform and SRE teams
Productionised a conversational health-monitoring agent that gives SRE and platform teams natural-language access to system health, sessions, and workload analytics on the HANA platform. It became the reference implementation for the wider fleet.
- Reference agent in production; other agents are built from it.
- Fleet under way across three domains: monitoring and alerting (triage, Dataproc and BigQuery monitoring, vulnerability tracking), operations (change coordination, patch-day automation, health analysis), and knowledge (L1 knowledge and troubleshooting agents).
LLM tool-use · LangGraph · RAG and vector stores · Embeddings · OpenTelemetry · n8n
In production
Real-time data platform at 200M+ events a day
Architected the Apache Spark and Dataproc platforms behind streaming and batch ingestion at multi-terabyte scale, with GKE-based orchestration and capacity monitoring.
- 200M+ events per day through PySpark pipelines over Pub/Sub streaming and batch ingestion.
- Large-scale GCP migrations delivered alongside, with the platform operated as code from day one.
PySpark · Dataproc · Pub/Sub · BigQuery · GKE · Terraform
In production
Self-service platform engineering and FinOps governance
Terraform provisioning frameworks and self-service tooling that removed most manual operations work, paired with a FinOps programme that made cloud spend a design constraint rather than a cleanup exercise.
- Eliminated about 80% of manual ClickOps through Terraform frameworks and GitOps workflows.
- Self-service platforms offloaded about 70% of L1/L2 tasks from the operations queue.
- Verified, recurring savings from BigQuery slot reservations, committed-use discounts, rightsizing, and quota management, with no performance regression.
Terraform · GKE · GitOps · BigQuery slots · GCP committed-use discounts · Rightsizing
