|
I build backends that companies actually run, and I measure computer-vision problems instead of wrapping another detector. About three years of that mix. Applying to MS programs. The production side is CRM, orders, attendance — schema, API, auth, UI, Docker, then someone uses it at work. The CV side is class confusion vs model scale, a C++ MOSSE lock on a detector box, and Raspberry Pi when the path has to leave the laptop. Focus
|
|
finegrained-yolo-scale — at which YOLO scale do visually similar categories stop being confused? mAP is not the claim. Primary metric is pairwise confusion on a held-out split. Sources are open-licensed only; the split is inside each source, so one photographer does not leak into test. After the box exists, a C++ MOSSE tracker has to lock, go Lost on a blank frame, and reacquire — a check that the box is usable in time, not only on a still. The n→s→m sweep is not finished. Empty cells in that table are empty on purpose.
|
People click these at work.
|
3D-print shop: printers, consumables, product / part / model tree, warehouse, shipments. In use at the company. Several client databases on one deploy, JWT / RBAC, React / TypeScript. |
|
QR check-in, report approval, scheduled backups. Construction defects: statuses, assignees, attachments. |
|
hh-assistant is a full product, not a script: embeddings for matching, cover-letter generation, Playwright, FastAPI dashboard, Telegram. LLM pipeline with a real UI, not a notebook. |
Coursework and edge hardware
Coursework, not a paper: PharmKursovaya — small DNN for glaucoma progression (binary classification, accuracy + AUC).
On a Raspberry Pi I wrote C for SPI / RSSI and an OpenCV preview path (rpiskanC). Radio + video on a weak ARM board, no cloud GPU in that loop.