👋🏽 Hi! I'm an M.S. student in Computer Science at UC San Diego, specializing in Robotics. I work in the Cognitive Robotics Lab, where my thesis is advised by Professor Henrik I. Christensen. I am interested in studying the intersection of machine learning and robotics, focusing on how large language models can make human-robot interaction more natural and useful.
Before that, I earned B.S. degrees in Mathematics-Computer Science and in Cognitive Science (specialization in Machine Learning and Neural Computation) at UC San Diego.
Email: cshukla [at] ucsd [dot] edu

Built Otto, an agentic AI system that monitors the robots' computer vision and insights pipeline for anomalies across 10+ BigQuery tables and debugs them with tool-calling sub-agents that query SQL, inspect code, and run shell commands. Deployed on GCP Cloud Run with Pulumi infrastructure-as-code.

Built a build-health quality review dashboard used by 80+ Apple teams, with a React.js frontend and a FastAPI/MongoDB backend. Integrated group-based OIDC authentication and managed Kubernetes deployments across staging and production.

Experimented with retail product recognition improvements by LoRA fine-tuning open-source vision-language models (+38% accuracy) and speeding up inference with vLLM. Designed a multimodal rank-fusion approach that reached 88% accuracy while running ~13x faster and cheaper than the existing method.

Designed a deterministic state machine to enable verification and E2E testing of negotiation workflow
A tour guide robot built from a Reachy Mini head on a ROSbot base, with RTAB-Map mapping, AprilTag landmarks, autonomous navigation, and conversation.
I had the opportunity to work as an Undergraduate Student Researcher under the guidance of Dr. Shan Lin on the following projects:
Springer Book Series: Algorithms for Intelligent Systems, A. Agrawal and C. Shukla

A five-part series building up autonomy on a small differential drive rover with ROS 2, from open-loop control to full area coverage.

EKF-based visual-inertial SLAM on Clearpath Jackal data from MIT's campus. Fuses IMU kinematics on SE(3) with stereo camera features to jointly estimate the robot trajectory and 3D landmark map, using sparse covariance slicing to keep updates fast. Also implemented Shi-Tomasi + Lucas-Kanade feature tracking.

SLAM for a differential-drive robot using wheel encoders, an IMU, a 2D LiDAR and an RGB-D camera. Dead-reckoned odometry is refined with ICP scan matching, then loop closures are enforced with GTSAM pose graph optimization to produce an occupancy grid and a floor texture map.

Estimated 3D orientation from IMU data by solving a constrained optimization over unit quaternions with projected gradient descent (PyTorch autograd), validated against VICON ground truth. The estimates were then used to stitch camera images into panoramas via equirectangular projection.

A Pomodoro timer robot on a MiniPupper 2.0 that keeps students focused by physically separating them from their phones. Using an OAK-D camera, AprilTags, and touch sensors, it detects when the user is distracted and responds with facial and audio emotes, driven by a ROS 2 state machine. I built the "device jail" and the emote audio. We ran a user study comparing it with a digital Pomodoro timer.



