Selected client work

A US technology services company runs a managed AI platform for its enterprise clients. Each client gets its own isolated deployment. The platform is two years old; I joined in June 2026.

Its agents answer questions and take actions across email, chat, call notes, campaign metrics and CRM data. One client's data must never reach another client's agents.

I own the agentic layer. I design the architecture and ship code with the engineering teams.

I decided which agent actions run on their own and which wait for human approval. Each tool an agent can call has its own permission scope.

I designed per-client data isolation. Retrieval scoping and tenant boundaries keep one client's context out of another client's prompts.

I defined how the platform connects to each client's systems: an events API and the authentication model.

Agents connect through two open protocols: MCP, which gives agents access to tools and data sources, and A2A, which lets agents from different systems work together.

The platform runs in production for a global enterprise client. Each new client gets its own agents on the same shared core, in an isolated deployment.

A pre-launch consumer AI product team needed an independent review of their retrieval-augmented assistant before opening a closed beta. The team had a working prototype and a launch timeline but limited confidence in retrieval quality, security posture, and regulatory readiness.

Reviewed the full system from ingestion through retrieval, ranking, generation, and response. Recommended a hybrid retrieval pipeline combining dense and BM25 search with reciprocal-rank fusion and cross-encoder reranking, a knowledge-base schema with consent and licensing metadata, threat modelling against agent and RAG-specific risks, reliability SLOs with stage budgets, and an evaluation harness with measurable launch acceptance bars.

Board-level review document and a phased four-month plan feeding the team's go/no-go process for closed beta.