Zongwei (Tom) Tang

Full Stack Developer Intern

• Automated post-session emotion analysis by building REST APIs that launch long-running inference jobs, track queued/running/failed/completed states, and return generated clinical reports without blocking user requests
• Reduced duplicate emotion-analysis jobs by implementing session-level idempotency keys, exponential-backoff retries, and request timeouts
• Built report-generation logic to aggregate model outputs into structured session results, validating malformed and incomplete inference responses before exposing them to the clinical application

React, FastAPI, Redis, Turso, Docker

• Full-stack research profiling app for PhD applicants, using the Semantic Scholar API to analyze professors' top-cited papers, publication trends, and frequent collaborators
• Moved long-running Semantic Scholar and LLM analysis into asynchronous background jobs with status polling, preventing request timeouts during multi-step report generation
• Built a multi-provider LLM router with token-bucket rate limiting, circuit-breaker failover, and background health checks to automatically route around unavailable providers
• Designed the data layer with SQLAlchemy, Turso, and Redis cache-aside invalidation, supporting professor, paper, and per-user search history
• Deployed with Docker Compose on a VPS, using Caddy as a reverse proxy and Cloudflare DNS-01 for automatic TLS certificate provisioning

OpenProject Club, UC Berkeley · NeRF, Open3D, NumPy, FastAPI, Plotly

• Built a pipeline converting real-world scenes into playable Minecraft worlds using NeRF, semantic segmentation, voxelization, and block conversion; owned the voxelization stage end-to-end
• Converted PLY point clouds with Open3D and NumPy, mapping 3D points to voxel grids, averaging per-voxel RGB, and resolving conflicting semantic labels by point density
• Deployed voxelization as a FastAPI microservice with endpoints for compressed NumPy output and interactive Plotly 3D visualization