• Led engineering on Mini Apps for Luma's AI image and video generation platform: one-click creative tools like Camera Angles and Relight, taken from prototype to paid launch in 1.5 weeks.
• Built out Luma's connector platform so creators move generated images and video between the AI canvas and Google Drive, Dropbox and Frame.io, via agent tools, streamed exports and one shared OAuth connect flow.
• Scoped and built custom MCP connectors end to end, letting users plug any remote MCP server into the AI agent: OAuth discovery with dynamic client registration, authless servers, and the stateless MCP wire format.
• Designed the MCP security model: DNS-rebinding-safe SSRF guard, Redis sliding-window rate limiting, and cross-server taint tracking against prompt injection.
• Integrated multilingual text-to-speech (Seed Audio) end to end, with 20 voices and usage billing on generated audio duration.

• Workstream Lead for the Facebook Lightweight Ad Creation interface, independently driving end-to-end product execution across ideation, scoping, PRD authoring, rapid prototyping and launch; used SQL analysis to measure experiment impact, validate hypotheses and inform roadmap priorities.
• Shipped Click-to-Messenger ad features that drove ~$2M in incremental ads revenue, directly increasing advertiser adoption.
• Delivered end-to-end web and mobile solutions with JavaScript, PHP, GraphQL and React, partnering with product, design, data science and ML teams to prioritize and ship the highest-impact experiments.
• Used AI coding agents (Claude Code and others) extensively to run 7–8 concurrent workstreams, greatly accelerating iteration across a large project portfolio.

• Built an internal Vue.js/Solr web app for searching and managing Points of Interest, adopted company-wide for map data operations.
• Ran user interviews and redesigned key workflows, cutting clicks for common tasks by 50%; added map layers with drawing and spatial search.

Joined a start-up that reduces food waste by partnering with restaurants and selling surplus food at discounted prices.

• Designed and built the help-center client and admin portals in Figma, React, Next.js and Node.js, used by 50+ restaurants including Subway and Jersey Mike's.

• Led the team structuring "Shareful", a resource-sharing mobile app, with MVVM architecture.
• Implemented UI components for Android and iOS using Kotlin, Jetpack Compose and Swift.
• Implemented backend features with Firebase.

• Developed an automated test suite to pinpoint defects, track bugs and validate functionality for POS software.
• Carried out the automated regression testing plan in JavaScript and prototyped automatic test-script generation, producing 20 automation test cases with trivial human intervention.
• Completed 68 manual tests and 20 automation test cases — 4,000+ lines of scripts and 80% test coverage.
• Investigated deviations in test results and determined root causes.

• Led business development and outsourcing, building partnerships with 10+ potential clients through game-industry conferences, business-matching systems and LinkedIn outreach.
• Proposed and facilitated B2B global marketing strategies for WeTest, Tencent's game-testing platform.
• Wrote weekly blogs about QA services and game testing, boosting public exposure by 300%.

• Held office hours for coding and project questions.
• Answered course questions on Piazza.
• Graded programming assignments and exams.

• Answered students' questions through Slack and email.
• Proofread course materials and managed them through Blackboard.
• Graded quizzes and exams on business operations.

• Led DTTDNet development, fine-tuning 3D localization to 85% tracking accuracy on a digital-twin RGB-D dataset captured with ZED2 cameras.
• Built a HoloLens 2 AR interface for Leo Rover's Martian exploration using Unity and Mixed Reality Toolkit.

Jeffrey Liu, Rajat Tandon, Uma Durairaj, Jiani Guo, Spencer Zahabizadeh, Sanjana Ilango, Jeremy Tang, Neelesh Gupta, Zoe Zhou, Jelena Mirkovic. 2022. Did your child get disturbed by an inappropriate advertisement on YouTube? In Proceedings of KDD Undergraduate Consortium (KDD-UC). ACM, New York, NY, USA, 7 pages.

Supervisor: Prof. Burcin Becerik-Gerber · Ph.D. student: Ala Nekouvaght Tak

• Researched reinforcement learning for personalized thermal comfort systems.
• Applied the Markov assumption to the RL agent's decision process and used a simple MLP to learn the policy given system states.
• Designed a custom reward function based on power consumption, thermal comfort and personalized control.
• Retrieved and managed intelligent-agent data using Python and InfluxDB.

Vice President of Finance (2021–2022) · Co-Chair of Professional Development (2020–2021)

• Connected with professionals from PwC, Deloitte and Boeing and hosted career-development events.
• Wrote budget plans; managed funding, accounts, membership fees and expenditures; hosted fundraisers.
• Designed and ran the official Ascend website.