Meta
Engineering at scale across full-stack product, machine learning, and generative-AI pipeline work. Public-safe summary — internal project names and metrics stay internal.
subham@earth:~$ ./wake --quiet
spatial memory loaded · camera detected · engine warm
Subham Banga
Building systems that understand and interact with the physical world.
Build log
SpatialMind
Where I'm heading in embodied AI: an open-source spatial-memory layer for AI agents — persistent, queryable maps of physical space, so an agent that has seen a room can reason about it later. Early-stage and unshipped — a problem being framed, not a product.
The academic spine
-
Selecting Source Tasks for Transfer Learning of Human Preferences
h. nemlekar · n. sivagnanadasan · s. banga · s. gupta · s. nikolaidis — icaros lab, usc
-
Sketch2Code: Transformation of Sketches to UI in Real-time Using Deep Neural Network
v. jain · p. agrawal · s. banga · r. kapoor · s. gulyani — arXiv:1910.08930 (2019) · widely cited
first prize — marconi society celestini program india (2019)
for Rakshak, an on-device speech-detection app for women's safety.
The roots half
sony a6700 · sigma 18–50mm f/2.8 · sony 70–350mm
$ cd ~/photography --all →Both halves
I build machines that understand space — and chase the light they can't compute.
I'm Subham — a software engineer at Meta working across full-stack, ML, and generative-AI pipelines, with a research background in human-robot interaction from USC's ICAROS Lab. I think in systems; I recharge where systems can't reach — behind a camera at first light, or on a motorcycle in the mountains. This site is both halves: circuits, with roots growing through them.
Open a channel
$ cat ./contact
> channels coming online