Selected work
Agentic Security
Microsoft · Project PerceptionDesigning the human-facing layer around agents that act on their own: how a person reads an unattended run, redirects it, undoes it, and stays accountable. Shipped as Project Perception, now in public preview inside Microsoft Defender. Half open — the launch film and what shipped are public. The case study behind it is password-protected; ask me and I'll hand the password over.
Docufy
Microsoft · SharePoint PremiumAn AI contract lifecycle platform built on SharePoint Premium's content AI services — metadata extraction, document classification, clause analysis. My design framework was adopted org-wide.
Passbook by Remitly
Time 100 Best Inventions '20An AI-assisted, tri-modal banking app — conversational chat, voice, and touch — built for immigrants navigating English-language and literacy barriers.
Frictionary
My own venture · MVP in pilotA product that runs a real AI agent through real tasks on your site and shows where it fails, and why. Designed, written and built solo — product, browser automation, data model and the deterministic detection layer behind the evidence claim.
There's more — Puente, Learner Management, and Smart RX — plus experience, skillset and additional impact.
Human control
Placed so agents can run more freely, not less
Reversibility
What can be undone decides what can be autonomous
Trust & red-teaming
Evaluating systems that don't repeat
Multimodal / VUI
Chat · voice · touch, one intent model
Agent-as-user
Designing for a visitor that isn't human
Microsoft · 2026—presentAgentic Security
Human control around autonomous security work — playbooks, sessions, legibility and reversibility — designed from research with 7 security personas.
Project Perception · public preview Microsoft · 2021—24Docufy
AI contract lifecycle on SharePoint Premium. Framework adopted org-wide.
3–6 months → under 14 days · Ignite 2023 Remitly · 2019—21Passbook
Tri-modal banking for immigrants navigating language and literacy barriers.
TIME 100 · 67% activation My own venture · 2026Frictionary
See your site the way an AI agent does — runs a real agent through real tasks and reports where it fails, and why. Designed and built solo, now in pilot.
$0.0115/run · 8.4s · MVP in pilot BuiltMCP workflows
Figma ↔ VS Code ↔ Copilot on one live component state.
~25% faster handoff BuiltSynthetic personas
AI-simulated testing of agent override behaviour.
~30% fewer live cycles BuiltFigma → Adaptive Card
AdaptiveLens — OpenAI + Gemini, emitting schema-valid JSON.
BuiltIntent coverage model
Mapped what users asked against what the assistant could answer, then instrumented the gap in Mixpanel.
+12% CSAT BuiltDeterministic detection
Friction found by code, not a model — so no finding can be hallucinated. A model only explains what code already proved.
evidence you can check