atomicrajat@edge

$ cd ~/about

About

I build intelligent systems that run locally, on the device — not as a replacement for the cloud, but as the right place for the work. Inference at the edge, coordination and heavy lifting upstream, and a considered line between the two.

My work sits at the intersection of edge AI, physical AI and robotics: making deep learning models actually run on power- and memory-constrained hardware. Optimisation, quantization, inference pipelines, and the unglamorous resource management that decides whether something stays a demo or becomes a product.

I work across edge boards and AI accelerators generally — the silicon changes, the problem doesn't. Fit the model to the compute budget, pick the precision you can live with, architect the pipeline so nothing sits idle. Currently that centres on NVIDIA's edge platforms, with simulation and ROS 2 for robotics.

Rajat M R
based in
Bangalore
projects
3
focus
physical AI
experience
4+ yrs

The next interesting phase of AI is physical: systems that perceive, decide and act in the real world, under real latency and power budgets. That's what I want to keep working on.

$ head -5 ./skills

NVIDIA JetsonEdge AI deploymentDeepStreamDeep LearningComputer Vision

$ cat ./experience4+ yrs · edge AI

  1. Jul 2026 — present

    Software AI Embedded Engineer / Solutions Architect

    TannaTechbiz — NVIDIA Partner Company · Bangalore

    • Design and deliver end-to-end edge-AI solutions on the NVIDIA stack for robotics, industrial automation and intelligent vision.
    • Build and optimise inference pipelines on Jetson with TensorRT, DeepStream and CUDA.
    • Ship real-time computer vision, multimodal and embedded AI onto edge devices.
    • Architect scalable, production-ready solutions with customers and cross-functional teams.
    • Tune models and deployments for performance, efficiency and reliability.
    TensorRTNVIDIA JetsonDeepStreamCUDAIsaac SimIsaac LabPipeline architecture

  2. Jan 2025 — Jun 2026

    Senior Software Engineer — Embedded Systems

    Vimaan Robotics India Pvt. Ltd. · Bangalore

    • Built the interface layer between sensors and applications using ROS.
    • Implemented algorithms to extend the existing Vimaan product line.
    • Led the firmware team on an end-to-end high-speed conveyor parcel-detection system, from design through implementation.
    Jetson XavierROSComputer visionProduct architectureDepth camerasLiDARProduction FW

  3. Jan 2024 — Dec 2024

    Embedded Firmware Engineer

    Vimaan Robotics India Pvt. Ltd. · Bangalore

    Developed the firmware stack on NVIDIA Jetson Xavier for interfacing cameras, depth cameras and LiDAR via ROS packages. Maintained production firmware releases.

    Jetson XavierROSDepth camerasLiDARProduction FW

  4. Aug 2022 — Dec 2023

    Edge AI Engineer

    SandLogic Technologies Pvt. Ltd. · Bangalore

    Built optimised edge-AI solutions across NVIDIA Jetson, Intel CPU, STM32, RISC-V and Cortex-M microcontrollers.

    JetsonIntel CPUSTM32RISC-VCortex-M

  5. Jan 2022 — Aug 2022

    Edge AI Intern

    SandLogic Technologies Pvt. Ltd. · Bangalore

    Conversion, optimisation and deployment of deep-learning models onto edge devices such as NVIDIA Jetson.

    Model conversionOptimisationDeployment

$ cat ./education

  1. Aug 2018 — Jul 2022

    B.E. Electronics and Communications

    KLE Technological University · Hubballi

    Coursework in Machine Learning, Basics of Embedded Intelligent Systems, Computer Communication Networks and Basics of Operating Systems.

    Machine LearningEmbedded Intelligent SystemsComputer NetworksOperating Systems

    cgpa 8.97

$ nvidia --boards

NVIDIA boards I work on — from an 8 GB Orin Nano up to Thor. Same problem at both ends: get a model that wants a data-centre to run on something you can hold. Drag either board to look around it.

orin-nanoglb · draco

scroll to load

Jetson Orin Nano Super

The board almost everything here runs on. 8 GB is a hard ceiling, which is exactly what makes it interesting — you cannot solve anything by throwing memory at it.

ai perf
67 TOPS
memory
8 GB LPDDR5
power
7–25 W
gpu
1024-core Ampere
agx-thorglb · draco

scroll to load

Jetson AGX Thor

The other end of the same line. Headroom to run the models I currently have to quantize, prune and thread around — and where the physical-AI work is heading.

ai perf
2070 TFLOPS
memory
128 GB LPDDR5X
power
40–130 W
gpu
Blackwell

jetson line

my range · Orin Nano → Thor

Nano
Orin Nano
Orin NX
AGX Orin
AGX Thor

Figures are NVIDIA's published headline numbers at different precisions, so they are not directly comparable — the rail shows the product line, not a benchmark.

3D models by Wendy Labs · visualisation assets, not official NVIDIA assets.

$ ls ./toolbelt

boards
Jetson Orin NanoJetson XavierJetson NanoRaspberry Pi PicoESP32K210
inference
TensorRTDeepStreamONNXINT8 / FP8 quantizationPruning
vision
PyTorchUltralytics / YOLOOpenCVMLXVLMs
robotics / sim
ROS 2Isaac SimIsaac LabIsaac ROSLiDARDepth cameras
on-device LLM
llama.cppwhisper.cppPiperGGUF / Q4_K_M
mcu / soc
STM32RISC-VCortex-MIntel CPU
hardware
I²C / PCA9685CSI & USB camerasServosKiCadCircuitPython
software
PythonTypeScriptNext.jsNodeSQLite

$ cat ./contact

LinkedIn is the quickest way to reach me — send a message and we can take it from there.

GitHubgithub.com/atomicrajat/atomicrajatLinkedInlinkedin.com/in/rajatmr/YouTubeyoutube.com/@atomicrajatEmailcontact@atomicrajat.in