ROS 2 Jazzy · RTAB-Map SLAM · Nav2 · Zenoh · Gazebo Harmonic · Python Mobly

• Designed and deployed an Autonomous Mobile Robot cluster in ROS 2 Jazzy to automate spatial customer user journey testing for Google Pixel hardware, replacing legacy testing rigs with dynamic 2D multi-agent testing across Find My Device, Bluetooth Channel Sounding, and GNSS validation.
• Built a navigation pipeline combining multi-sensor EKF state estimation (LiDAR ICP, RealSense D435i depth/IMU, wheel odometry) with RTAB-Map graph SLAM, surveyed AprilTag landmarks, and Nav2, reaching 5 cm waypoint precision on multi-hour runs.
• Scaled multi-robot fleet coordination over corporate Wi-Fi by migrating from DDS multicast to a distributed Zenoh cluster (TCP unicast & POSIX shared memory), cutting idle traffic to 0 KB/s, capping telemetry under 300 KB/s, and synchronizing clocks via Chrony NTP for sub-millisecond alignment.
• Streamlined autonomy validation and sim-to-real parity by using a 1:1 Gazebo Harmonic digital twin with custom URDF kinematics, sensor simulation, and physics plugins to test navigation stacks virtually prior to deployment.
• Built a decoupled Python client library and Mobly test framework communicating over WebSockets, enabling corporate test harnesses to seamlessly deploy and execute automated test jobs on physical rovers over the cloud with zero ROS dependencies in depot.

PyTorch · Flower · Federated Learning · Multimodal Transformers · Gemini Live · VLMs

• Architected a federated learning pipeline using PyTorch and Flower to fine-tune an adaptive multimodal fusion model ensembling multiple unimodal transformer models (facial and acoustic representations) for continuous Valence-Arousal tracking, dynamically driving Gemini Live conversational intervention timing.
• Developed a real-time assistive vision pipeline querying a backend Vision-Language Model (VLM); implemented a token margin uncertainty quantification algorithm on log-probabilities to dynamically modulate spoken dialogue confidence.

ROS 2 · Unitree L2 LiDAR · Point-LIO · RTAB-Map · STVL · Nav2 MPPI · NVIDIA Isaac Sim

• Leading a team of 6 engineers architecting the autonomous navigation stack in ROS 2 for a multi-terrain rover competing in the University Rover Challenge (URC).
• Engineered a LiDAR-inertial SLAM pipeline fusing Unitree L2 LiDAR and IMU data via Point-LIO into RTAB-Map, integrating a Spatio-Temporal Voxel Layer (STVL) for 3D point cloud obstacle persistence and dynamic 2D/3D local costmap generation across GNSS-degraded environments.
• Validated end-to-end autonomy in NVIDIA Isaac Sim by modeling custom URDF kinematics, simulating raw LiDAR/RGB-D sensor streams, and tuning the Nav2 MPPI controller over rough terrain for real-time dynamic trajectory optimization.

U.S. Patent Pending · Malicious AI Model Detection · Inline Firewall Blocking · Gemini API · GCP Cloud SQL

• Architected a data curation pipeline where the core LLM classifier (built with Gemini API) achieved 100% precision/recall, creating a high-quality ground truth dataset managed with GCP Cloud SQL and Cloud Storage.
• Engineered a production firewall feature for inline, real-time blocking of malicious AI models, achieving 100% detection coverage of malicious samples with zero false positives across 7000 known samples.
• Developed and optimized the core detection algorithm for high-throughput, low-latency performance in resource-constrained firewall environments; utilized a GCP VM as a secure sandbox for all development/testing.
• Extended the internship to continue work on reverse-engineering malware families using YARA rules and static/dynamic analysis.

Python · TF-IDF · DBSCAN · Rule Mining · CI/CD Pipeline

• Developed an advanced threat intelligence system (Python) that automates identifying and deploying production firewall rules for new malware families using ML clustering (TF-IDF, DBSCAN) and association rule mining to process millions of logs, slashing rule creation time from days to minutes.
• Established a robust CI/CD pipeline with automated testing to accelerate development.