point cloud · sample_office_scene.ply · 120k pts
This is a subsampled preview (120k of ~2M points). Watch the full reconstruction video ↗, or browse the full playlist ↗.
diwakar@robotics — zsh click to restore ↑

$ whoami

Hi, I'm Diwakar Ravichandran

Robotics Engineer & Computer Vision Researcher

M.S. in Robotics from UC Riverside, specializing in SLAM, computer vision, and autonomous systems. I take research ideas all the way to shipped, production code — from drone-based 3D reconstruction to real-time, GPU-accelerated perception pipelines.

View My Work Download CV
Diwakar Ravichandran ● tracking · pose locked
02 — about

About Me

I graduated with a Master's degree in Robotics from the University of California, Riverside, where I specialized in computer vision, SLAM (Simultaneous Localization and Mapping), and autonomous systems. My thesis, Celesta, is a fully differentiable optimization framework that integrates distributed bundle adjustment with Leiden-based graph partitioning for scalable, GPU-accelerated visual SLAM using NVIDIA Thrust.

Previously, I worked as a Data Scientist at Jio Platforms Ltd., where I developed and shipped computer vision solutions for drone-based tower reconstruction using SLAM. I contributed to video analytics for surveillance and visual document understanding, delivering production-ready pipelines from prototype to deployment.

My technical toolkit spans C++, Python, MATLAB, and ROS, with hands-on experience in GPU programming (CUDA) and NVIDIA platforms. I care about writing maintainable, performant code and bringing research ideas into deployed systems — and I'm always chasing the next hard problem in robotics and AI.

M.S.

Robotics Degree

5+

Years Experience

10+

Research Projects

Diwakar at graduation
03 — projects

Featured Projects

Fused attention kernel: N×N score matrix, online softmax, KV-cache compression, W4A16 matmul

LLM Inference Kernels

Custom CUDA kernels for LLM inference — fused FlashAttention-style attention, INT8/INT4 KV-cache compression, and W4A16 quantized matmul — profiled with Nsight and benchmarked against PyTorch, vLLM, and FlashAttention on Llama 3 8B.

CUDA LLM Nsight
Dense 3D Reconstruction Engine

Dense 3D Reconstruction Engine

Helper repository to visualize dense 3D reconstructions from SLAM and structure-from-motion pipelines.

Python 3D
Celesta

Celesta

Dockerized demo of Celesta — distributed, GPU-accelerated bundle adjustment built on Facebook's DABA with improved load balancing across GPUs. Explore the repository for implementation details and usage.

CUDA Optimization Thrust Docker
Bundle Adjustment

Bundle Adjustment

GPU-accelerated bundle adjustment for structure-from-motion and SLAM using CUDA for real-time optimization.

CUDA C++ SLAM
GNSS/INS EKF estimated trajectory vs ground truth on KITTI

GNSS / INS Extended Kalman Filter

A loosely-coupled 15-state Extended Kalman Filter fusing GNSS position fixes with inertial measurements on the KITTI raw dataset, including a GPS-dropout dead-reckoning demo. Implemented in NumPy/SciPy.

Python EKF Sensor Fusion KITTI
Photo evaluation agent pipeline: facial analysis, aesthetic scoring, CLIP/BLIP, DINOv2+FAISS, DETR/ViT

Photo Evaluation Vision Agent

A computer-vision agent that analyzes photos end-to-end — facial analysis, aesthetic scoring, CLIP/BLIP semantic tagging, DINOv2 + FAISS similarity search, and DETR/ViT scene understanding.

Python CLIP DINOv2 FAISS
Collaborative V2V Communication System

Collaborative V2V Communication System

Vehicle-to-vehicle communication systems for intelligent collaborative driving. Autonomous vehicles in CARLA simulation with multi-agent coordination.

Python CARLA Autonomous Driving
Vision Transformer: image patches to token sequence with CLS and positional encoding into an N-layer encoder

MNIST-Former

A compact Vision Transformer trained on Fashion-MNIST, with training and inference separated from visualization and optional Nsight profiling hooks for GPU timeline analysis.

PyTorch Transformer ViT
Semicon Bayesian Drift

Semicon Bayesian Drift

Bayesian drift modeling and analysis for semiconductor applications. Jupyter-based workflows for inference and visualization.

Python Jupyter Bayesian
MNIST with C++ and CUDA

MNIST with C++ and CUDA

A fun project implementing a simple neural network in CUDA for MNIST digit classification, written in C++ with GPU acceleration.

C++ CUDA ML
Two-link robot arm with joint angles theta-1 and theta-2 reaching a target along a planned path

Foundations of Robotics

A ROS playground for forward/inverse kinematics, open- and closed-loop control, and 2D path planning, visualized in Gazebo. From UC Riverside's EE283A (Foundations of Robotics).

ROS Gazebo Control
04 — research

Research

Celesta

2024

A fully differentiable optimization framework that integrates Distributed Accelerated Bundle Adjustment (DABA) with the Leiden algorithm for improved graph partitioning in visual SLAM. Implemented with NVIDIA Thrust, Celesta achieves scalable, GPU-accelerated bundle adjustment with balanced workloads and better convergence than Louvain-based partitioning. Master's thesis, UC Riverside.

SLAM CUDA Optimization
Download thesis (PDF)
05 — skills

Skills & Technologies

Programming & Robotics

Python
C++
ROS
MATLAB ®

Computer Vision & ML

Image Processing
Video Processing
SLAM
Deep Learning

Tools & Frameworks

PyTorch
CUDA ®
OpenCV
TensorFlow

Shipping & Deployment

Docker
FastAPI
Kubernetes
CI/CD
06 — contact

Get In Touch

Let's build something together.

I'm always interested in new opportunities and exciting problems in robotics, perception, and AI. Whether you have a question or just want to say hi, feel free to reach out.

diwakarjravi@gmail.com
UC Riverside — Robotics
Fremont, CA, USA