AEGIS - VBS 2027 Video Search
Built a live-first VBS 2027 retrieval system with a 6-member RMIT team. Led performance benchmarking and the public project page.
- Role
- Performance & Benchmark Engineer
- Timeline
- 2026
- Team
- 6 Members
Highlights
- Parallelized Type-2 VQA crop scoring - cut 75-220s down to 15-30s
- Added system-wide fail-closed error handling - live search never goes down
- Fixed the eval suite's miscounting - benchmarks now report real numbers
- Built the team's public project page in React + Vite - the team got its own landing page
Technology
- Python
- FastAPI / Uvicorn
- React / Vite
- TypeScript
- Tailwind CSS
- shadcn/ui / Radix
- Lucide
- Qdrant
- PyTorch / Transformers
- vLLM
- OpenAI VLM
- WeMM Embedding
- SigLIP
- YOLOE / Ultralytics
- PaddleOCR
- Whisper
- OpenCV / Pillow
- BM25 / RRF
- DRES API