Youran Ma

• Built an AI interviewer product in an Agile team, translating customer requirements into sprint-ready tasks and coordinating PR reviews and merges to deliver backend and frontend features on schedule.
• Improved interview-flow reliability by adding recording duration and silence-detection checks in mock sessions, refining prompts and retry guidance to reduce invalid recordings caused by incorrect user actions.
• Implemented sentence segmentation for the speech system to deliver smoother, low-latency chunked TTS streaming.
• Engineered a scalable recruitment dashboard with a Vue 3 frontend and backend APIs for paginated lists, summary cards, and top-ranked views; added filtering and sticky headers to reduce full-data transfer and improve rendering performance.

• Migrated a legacy PHP-based workflow into a Qt/C++ PDA application supporting barcode scanning, material receiving, and inventory location transfers, improving usability for shop-floor operations.
• Modularized request and response handling by encapsulating Qt native JSON classes into reusable components, accelerating feature development and improving team productivity.
• Assisted in the Java and Spring Boot backend, adding business logic and database integration for frontend features.
• Worked with production teams to refine workflow details and supported testing, deployment, and release handoff in the company Alibaba Cloud environment.

Stack Vue 3, Vite, Rails, PostgreSQL, GitHub Actions, Heroku

• Worked in a Scrum team of four as Tech Lead, coordinating sprint planning, task breakdown, and integration between a Rails RESTful API backend and a Vue 3 plus Vite SPA.
• Designed the PostgreSQL data model and Rails MVC endpoints to power scalable team-composition discovery and browsing, adding indexed server-side filtering and pagination to prevent UI freezes as the dataset grew.
• Led the Vue 3 frontend, implementing composition search, editing, and recommendation features while improving accessibility with semantic markup, ARIA, and keyboard navigation.
• Contributed to CI/CD through GitHub Actions running RSpec and Cucumber, then deployed to Heroku with Postgres, Google OAuth authentication, and SMTP email delivery.

Stack C++, CUDA, PyTorch C++ Extensions, Python

• Implemented the forward pass of scaled dot-product attention using PyTorch C++ extensions and CUDA, building naive, tiled, and fused Softmax plus P@V kernels with validation and benchmarking scripts.
• Optimized the two major computation stages with shared-memory tiling and warp-level reduction, and fused Softmax plus P@V to avoid writing probability matrices back to global memory.
• Built correctness and performance evaluation pipelines in Python and benchmarked sequence lengths from 64 to 1024 with head dimensions from 32 to 256 on an RTX 2060 SUPER, achieving about 3x speedup over a PyTorch CPU baseline.