"Because apparently running neural networks on hardware that barely has enough RAM is considered fun."
I'm an electronics and robotics engineer who enjoys making hardware do things it was never designed to do.
Most people deploy AI on GPUs. I deploy it on microcontrollers, then spend three days wondering why I have 12 bytes of SRAM left.
My work orbits Embedded AI, Robotics, FPGA acceleration, TinyML, Edge Vision, and Distributed Embedded Systems — where the goal is simple:
Make hardware think before it runs out of memory.
class Sumit:
currently_breaking = ["STM32", "ESP32", "FPGA", "Raspberry Pi"]
research = [
"Embedded AI",
"On-device Learning",
"Federated TinyML",
"FPGA Accelerators",
"Computer Vision",
]
hobbies = [
"Reading 1200-page datasheets",
"Optimizing until the compiler gives up",
"Removing printf() to save Flash",
"Finding one missing semicolon after 3 hours",
]
current_status = "Compiling..."|
⚡ Embedded Systems 
📡 Communication 
|
🧠 AI / Edge AI 
💻 Languages 
|
|
Federated learning on an STM32, because cloud servers are expensive.
|
Teaching an ESP32 to learn new patterns without forgetting the old ones. Humans should try this too.
|
|
Real-time monocular 3D understanding on hardware that definitely shouldn't be doing 3D reconstruction.
|
Seeing how much AI fits into programmable logic before the timing reports start crying.
|
|
Secure telemetry, because sending plain-text packets in 2026 feels illegal.
|
Motion detection without throwing a neural network at every engineering problem. Mathematics still works.
|
Currently exploring:
- 🧠 On-device Learning
- ⚡ FPGA AI Acceleration
- 🤖 TinyML & Edge AI
- 📡 Federated Learning
- 👁️ Embedded Computer Vision
- 🔋 Resource-Constrained Intelligence
I maintain an ORCID researcher profile documenting my published and ongoing research.
- 🧩 Build AI that doesn't need a GPU.
- ⚡ Make microcontrollers learn on their own.
- 🔧 Design hardware that survives outside the lab.
- 🤖 Build robots that don't need cloud permission to think.
- 📖 Reduce the number of tabs containing datasheets. (Currently failing.)
Research Papers ████████░░ 80%
Firmware Bugs ██████████ ∞
Coffee ██████████ 100%
SRAM Available ██░░░░░░░░ 12 bytes
Datasheets Open ██████████ 27 tabs
Sleep ░░░░░░░░░░ optional
while (true) {
Learn();
Build();
Break();
Debug();
Repeat();
}"If it compiles on the first try, I'm immediately suspicious."
⚡ Always one compiler warning away from another project.