I build machines that have to work out where they are and what they are looking at — a racing car with no driver, a drone mapping a mine where GPS cannot reach, a camera sorting waste on a moving belt. Most of my useful work has been figuring out why something failed when the obvious explanation was wrong.

class ShivenduKumar:
    def __init__(self):
        self.education   = "B.Tech Mechanical Engineering @ NIT Patna"
        self.cgpa        = 8.06
        self.graduating  = 2027
        self.fellowship  = "Chanakya UG Fellow, TEXMiN @ IIT (ISM) Dhanbad"
        self.papers      = 2          # one as first author
        self.available   = "6-month internship from mid-December 2026"

    def what_i_actually_do(self):
        return {
            "perception":  ["sensor fusion", "SLAM", "vision transformers"],
            "control":     ["PID", "Pure Pursuit", "unscented Kalman filter"],
            "simulation":  ["ANSYS CFX", "LES-WALE", "FW-H acoustics"],
            "when_stuck":  "stop assuming, go look at the raw data",
        }

Two ways of finding people: YOLOv5 on the surface, simulated GPR below it. 84.7% detection confidence across 50+ runs. The part worth talking about is not the number — it is knowing which conditions made each sensor lie.

A* global planner with dynamic obstacle avoidance and cascaded PID control, built in Webots.

Python PyTorch Webots SciPy NumPy

📄 Published — Modeling and Simulation of UAV-Based Search and Rescue, ICAMAS 2026

→ Repository

Nobody assigned this one. A full autonomous racing stack on Linux:

• 2D LiDAR SLAM with particle-filter (AMCL) localisation
• Closed-loop Pure Pursuit path tracking
• Reactive gap-finding controller that picks a line through obstacles at speed

ROS C++ Python Gazebo Linux

→ Repository

Underground tunnels defeat visual localisation: no ambient light, and brick walls that repeat every few metres so every frame looks like the last.

An unscented Kalman filter fusing 100 Hz IMU with sparse visual features and LiDAR cut positional drift 82.4%, down to 0.07 m over a 60 m gallery. Artificial potential field for obstacle avoidance, and a 3D digital twin built from the LiDAR returns.

ROS2 Gazebo Fortress UKF V-SLAM

📄 First author — AI-Enabled Autonomous Mapping in Underground Environments · Funded by the TEXMiN Chanakya Undergraduate Fellowship

Owl-inspired serrations on a UAV propeller, three geometries, transient large-eddy simulation with the Ffowcs Williams–Hawkings acoustic analogy.

• Full-span serration cut noise by 3.87 dB (OASPL)
• Serrating only half the blade made it 1.87 dB worse
• The cause was not the serration but the junction between treated and untreated sections

The transition was wrong, not either half. That result is the reason the study was worth doing.

ANSYS Fluent LES-WALE FW-H Fusion 360

→ Repository

A 500 × 500 × 500 mm chassis carrying 20 kg at 1 m/s, drivetrain sized from first principles to a 2.0 safety factor (127.3 RPM, 27.5 N). 6 mm aluminium 6061 base plate, an ABS mezzanine isolating the compute from motor vibration, and a TS35 DIN rail so standard components mount without custom brackets.

Fusion 360 Kinematics Structural analysis

Semifinalist from 115,000+ registrations. A healthcare analytics product taken from research to working software in a competition sprint — market and competitor research, then requirements and wireframes, then a multi-agent pipeline with an NLP layer turning unstructured reports into ranked KPI dashboards a non-technical reader can act on.

Multi-agent AI NLP Streamlit Figma Python

→ Repository

mindmap
  root((Shivendu))
    Perception
      Sensor fusion
      SLAM and localisation
      Vision under degraded input
    Control
      Unscented Kalman filtering
      Closed-loop tracking
      Path planning
    Simulation
      LES and aeroacoustics
      Turbomachinery CFD
      Digital twins
    Open questions
      Why vision fails when input degrades
      Localising without GPS or light

Open to a six-month internship from mid-December 2026

Robotics · computer vision · simulation · product

kumarshivendu47@gmail.com