From hackathon POC to leadership buy-in

Built POCs that made real-time coauthoring user impact visible and surfaced the technical challenges transparently. Leadership hesitated on signing a large investment — the POC turned an abstract pitch into a concrete, testable bet. Made the cost, risk, and payoff tangible, which secured the buy-in.

Adapting when the default doesn’t fit

The standard Fluid protocol assumed native .fluid files — but SharePoint Pages stores data in .aspx list items, a fundamentally different model. Led the design of a derived architecture that diverged from the default Fluid path, requiring cross-team alignment with Fluid APIs and Push teams to agree on a new storage and sync contract. Drove it to completion as tech lead.

Designing for safe contribution at scale

Designed the architecture with schema constraints and validation so non-senior engineers could safely contribute features without risking data integrity. Invested in reusable, perf-optimized components (e.g. presence indicators) that encapsulated complexity behind clean APIs.

Slowed down when it matters

Deliberately slowed rollout velocity to avoid data loss. Spent 3 months building a dedicated data-loss dashboard, fixing data-loss related issues, and adding cross-team observability — aligning Push and Fluid APIs metrics into a unified dashboard across 3 teams. Ensured we could detect and resolve problems end-to-end before scaling to broader audiences.

Swimlane and delegation

Cross-team collaboration requires clear problem statements, proposed alternatives, and efficient communication. Delegation isn’t just about reducing load — it’s about enabling the team to scale faster, giving others growth opportunities, and mentoring junior engineers into independent owners of their areas.

Designed the AI-first abstraction layer

SharePoint Pages had no LLM-readable contract — components were opaque blobs. Proposed and led the design of an AI-first schema that distilled complex page data into structured JSON the model could reason about. Reduced token usage by >50% and significantly improved model focus — becoming the foundation for every AI feature that followed.

Scoped down to ship through ambiguity

Full-page AI editing was undefined — unclear scope, unproven value. Defined a scoped-down plan targeting section generation: bounded risk, measurable customer value, and fast iteration cycles. Shipped Smart Sections to GA (~60K MAU), validated assumptions with real users, and built the infrastructure that scaled to full-page later.

Turned nondeterministic into measurable

LLM output is inherently unpredictable. Built a multi-channel evaluation framework — assertion-based checks, LLM-as-judge scoring, and visual critic analysis — to evaluate content quality comprehensively. Gave the team a repeatable way to measure what was previously unmeasurable.

Identified foundational needs, executed fast

Recognized AI Properties as the critical enabler before any feature could ship. Built and rapidly iterated a POC and design doc that proved the value proposition — leading to adoption by multiple partner feature teams across the organization.

Navigating constant trade-offs

Every decision involved balancing latency, content quality, and system complexity. Developed the judgment to navigate these trade-offs quickly — knowing when to accept higher latency for better output, when to simplify the architecture, and when good enough is the right call.

Engineering excellence with AI in the loop

Embraced AI tooling for velocity while maintaining critical thinking and owning final judgment. Helped the team understand hallucination boundaries, build intuition for when to trust and when to verify, and keep engineering rigor as the bar — not speed alone.