Experience · Microsoft · April 2021 – 2026

From product engineering to the quality of AI systems.

I joined Microsoft in April 2021. My work moved from product engineering, through AI-powered product experiences, to measuring the quality of AI responses and tuning agents with expert know-how.

Chapter 01 · April 2021 – September 2023

Product and Platform Engineering

Product engineering on Microsoft Feed, from shipping features to owning them in production. This is where I learned that measurement and reliability are part of the product.

Microsoft Feed in Outlook mobile

Key contributions

Shipped SharePoint News Reactions and brought social signals into the feed.
Built dwell-time engagement measurement and fixed lost telemetry on mobile.
Ran compression experiments; response compression rolled out after a controlled experiment.
Fixed keyboard focus visibility, and supported the feed through on-call and incident investigation.

01 Professional

Building and Improving Production Feed Experiences

Microsoft Feed brought relevant work content into Microsoft apps. I shipped reactions for Viva Engage and SharePoint News, built dwell-time measurement and made the feed load faster.

Microsoft · 2021–2023Read case study

Chapter 02 · Late 2023 – 2026

AI-Powered Product Experiences

My work moved from conventional product engineering to AI-powered experiences: first features that summarise and surface information in existing products, then experiences built on agent workflows. The client now had to request work, follow its state, stream results and fail clearly. Not all of these experiences were released publicly.

Copilot in Outlook · Catch-up experiences in Microsoft 365 Copilot · A daily-briefing agent

Key contributions

Moved catch-up onto an agent architecture, with gateway routing, client changes and meaningful errors when the AI service failed.
Built the meetings and follow-ups scenario in the AI processing pipeline, including citations.
Took a proactive daily-briefing agent through packaging and store validation to its internal launch.
Traced an authentication failure after a workflow migration to the wrong kind of token, and with a colleague fixed false error messages shown when a completed answer existed.
Added the first monitors and troubleshooting guides, and consolidated prompt snapshot tests to cover every prompt type.

01 Professional

Catch-Up Experiences in Microsoft 365 Copilot: From Feature to Agent

Helping people catch up on the meetings, follow-ups and messages they missed. I worked on grounding and reliability, moved the experience onto an agent architecture, and built the monitoring and tooling that kept it healthy.

Microsoft · 2023–2025Read case study

Chapter 03 · Primarily 2025 – 2026

AI Quality, Evaluation and Reliability

Measuring whether AI-generated responses are good, finding out why they fall short, and making the sources behind them reliable. Two different problems, each with its own case study.

Researcher and Copilot Citations in Microsoft 365 Copilot

Key contributions

Developed the claim density metric, building on ideas shared with a research colleague, and built a temporal relevance metric, both integrated into offline evaluation.
Led regular baseline runs to separate real changes from noise.
Investigated context and grounding failures, and implemented a change to preserve context between workflow stages.
Investigated why citation markers leaked and shipped detection for them, helped move citation grouping into the service, and made citation problems measurable without collecting user content.

01 Professional

Researcher: Measuring and Improving AI Response Quality

Researcher produces grounded reports in Microsoft 365 Copilot. I built evaluation metrics, ran offline baselines and experiments, and investigated why responses lose context or fall short.

Microsoft · 2025–2026Read case study 02 Professional

Copilot Citations: Improving Response Reliability

Citations let people check where an AI answer came from. I investigated why citation markers leaked and shipped detection for them, helped move citation grouping into the service, and made citation problems measurable without collecting user content.

Microsoft · 2025–2026Read case study

Chapter 04 · 2026

Microsoft 365 Copilot Tuning

Helping subject-matter experts capture how they work and turn that knowledge into reusable agent playbooks, instructions and quality checks. I owned the skill-suggestion experience.

Workflow discovery and skill suggestions

Key contributions

Integrated workflow-discovery and skill suggestions into the expert interview and playbook workflow.
Worked on letting builders accept and refine suggested skills through the Work IQ connector.
Helped decide when AI-assisted refinement happens, keeping expert review and builder approval distinct.

01 Professional

Copilot Tuning: Capturing Expert Know-How to Tune AI Agents

Helping experts turn how they work into agent playbooks. I owned the skill-suggestion experience, which surfaces likely workflows from Microsoft 365 activity so experts don't start from a blank page.

Microsoft · 2026Read case study

Internal systems, customer information and unpublished figures are left out for confidentiality.