Leading the innovation team.

Teaching first

Before we built anything, I designed and ran an AI curriculum for the team: how LLMs actually work, where they fail, and how to apply them to daily engineering and to product features. Then we picked the places AI pays off most, product creation and audit in the company's marketplace, and shipped there.

These are company products, so the source is private and the diagrams are simplified.

AI Product Search

Text, voice and image queries through one hybrid keyword and vector pipeline, with runtime-switchable model providers.

• Typed, spoken and photographed queries run through one pipeline: OpenSearch hybrid retrieval blending BM25 keyword scores with vector search.
• An LLM turns natural language into filters, but it only sees brands and categories that exist in the catalog. Providers switch at runtime between Bedrock, OpenAI and Gemini.
• Keyword search runs first and the LLM steps in only on zero results. Rate limits and model outages fall back to keyword.
• Merchandisers get six placement strategies for promoted products, plus synonyms, redirects and an A/B search-test console.

Compliance Analyzer

Contracts checked clause by clause against configurable rule sets. Every flag cites its clause and a human signs off.

• Reads contracts, policies and agreements clause by clause against configurable rule sets, and explains every deviation it flags.
• Prepares the review in minutes instead of hours of reading, while the final judgement stays with the reviewer.
• Every finding is grounded in the supplied documents and cites its source clause. Nothing is auto-approved.

Order-placement MCP server

A Model Context Protocol server over SAP Commerce APIs, so an AI agent places orders through a few safe tools.

• A Model Context Protocol server over existing SAP Commerce APIs, so an AI agent places orders through a few intent-level tools instead of dozens of raw endpoints.
• Every response is filtered before it reaches the model. Credentials and privileged calls stay inside the MCP boundary.
• SAP keeps its native business rules and security model while the ordering flow becomes agent-driven.

AI Travel Assistant

Plain-language chat in, an end-to-end itinerary with live flight options out. Prompts tuned without redeploys.

• A conversational assistant that produces end-to-end trip itineraries and live flight options from natural-language chat.
• Models and prompts are managed from a React admin console, so the business retunes behaviour with no redeploys.
• Flight and itinerary data from provider APIs is normalised into structured, agent-ready responses.