Project Author
A desktop agent purpose-built for SJTU students: through public campus APIs it brings Canvas, university
email, course schedules, lecture replays, and the campus forum into a single AI assistant, wrapped as
tools and skills the agent can read and write directly — solving the pain of scattered learning
information and hard-to-reach resources.
-
Ownership from 0 to 1:
deeply involved through the full incubation — original idea, product design, and the demo — then
assembled the team, planned development, and coordinated promotion.
-
Remote projects for server-side development:
over SSH, Dela uploads the compiled dela-helper runtime binary to the server and keeps a stable,
high-throughput SFTP channel between server and local machine; path-prefix rewriting and remote
tool-call optimizations let the agent read/write server files and run Bash commands as if local, with
file changes synced in real time.
-
Local-first architecture & credential safety:
an Electron + React + TypeScript app on a Pi-based agent kernel, compatible with MCP, third-party
skills, and most of the Pi Packages ecosystem; credentials are encrypted in the system keychain and data
never leaves the local machine.
Core Developer
A cross-platform toolbox aimed at media and video creators, integrating media downloading and video
transcoding to solve the problem of scattered creation tools. Built on Tauri; the release installer has
been downloaded ~3,000 times.
-
Decoupling dependencies from the app:
moved from the bundled Full/Lite dual-edition installers to a unified installer with on-demand install
and independent updates via a China CDN — dependency updates and app updates ship separately, cutting
repeated downloads and distribution bandwidth while eliminating dual-edition maintenance.
-
Unified task management:
designed a tool adapter layer that converts tool-specific parameters and execution into a unified task
description; a shared task queue handles queuing, concurrency control, status, and progress feedback, so
new tools reuse the existing pipeline out of the box.
-
End-to-end product ownership:
independently drove feature development, build/test CI/CD, website design and deployment, and Microsoft
Store release and promotion; GitHub Releases auto-publish to the Microsoft Store via CI and go live once
approved.
Temporal Awareness and Adaptive Waiting for GUI Agents (SJTU 49th PRP)
2026.03 – 2026.09
Experiments & analysis · 3-student research team
Studied temporal awareness for GUI agents on the OSWorld benchmark with DART GUI 7B: fixed waits make up
~84% of a task's runtime, and the project explores letting the model decide wait durations on its own.
-
Wait-strategy experiments & pipeline speed-up:
screened a stable 67-task subset via three full-score runs to quantify the trade-off between wait
length, task performance, and interaction count; located and removed hidden pauses, cutting the
action-request round-trip from ~124 ms to 40 ms (-68%).
-
Wait labels & LoRA post-training:
built a wait-label dataset from post-action screenshot sequences with manual review; the post-trained
model reaches 95.1% bucket match on seen data (48.5% unseen), identifying poor long-wait generalization
as the key blocker to real speed-ups.