i write deep-learning kernels (layernorm, softmax, etc.) for TI's DL library, targeting inference on TDAx edge accelerators. on the side, i fine-tune small VLAs for warehouse pick-and-place on a single 5070 Ti.
day job and side project are the same shape: making small models do real work on constrained hardware. posts here are experiment writeups and negative results from the side-project half.
ACT → SmolVLA → harder scene on one 5070 Ti. Peaked at 60% on a 4-color shelf, dropped to 17.5% when I shuffled the books (0/54 on non-canonical positions). The textbook fix made it strictly worse, and the harder scene went 0/10. Three architectures, four scenes, zero clean wins. Full log.
Notes on resuming my undergrad thesis — single-image 3D reconstruction with a text prompt — as a web app.