Kiarash Zamani

Built an on-device multimodal inference pipeline on a Jetson Orin Nano that indexes camera frames into a FAISS vector database, so you can search them by description.

Used normalized CLIP embeddings to pick the top-k relevant frames before anything reaches the vision-language model, which keeps most of the expensive VLM inference from running at all.

Added incremental embedding and a persistent index, plus periodic sampling and deduplication to keep storage down. Covered it with a 32-test suite that mocks out the model and hardware.

Trained a YOLOv8 defect detector on a 1,000+ image dataset for aerial power line inspection, reaching 82%+ mAP with augmentation and contrast normalization.

Deployed an asynchronous FastAPI inference service with endpoint-level reporting for average / p95 latency and throughput.

Load tested it from 1 to 25 concurrent requests and brought p95 latency down with headless inference, holding about 5 req/s on CPU. Built a reproducible MLOps pipeline with automated preprocessing and artifact versioning so training and API output stay in sync.