I build high-fidelity digital twins of real intersections so that 3D perception models — trained entirely in simulation — work the moment they're deployed on a real LiDAR. Closing the Sim2Real gap is my dissertation, my postdoc, and most of my GitHub.
I'm a researcher who fell in love with computer science through game dev as a teenager, kept the obsession through three degrees, and now spend my days at the intersection of simulation, perception, and the real world.
My Ph.D. (UCF, 2025) — From Failure to Fidelity: Enabling Scalable Sim2Real LiDAR Perception — argued that the gap between simulated and real LiDAR is mostly a data problem, not a model problem. The work has since produced an open-source framework, three public datasets, a JOSS paper, and a pending patent.
Outside the lab I build things for fun: a VR narrative game for Quest 3, a voice-to-keyboard dictation tool I actually use every day, a personal AI assistant with an Android frontend. Eight years of programming, four languages I'm fluent in, and one weakness for over-engineering side projects.
One thesis, four threads. They all share a question: how do we get a perception model to behave the same way on real sensor data that it does in a simulator?
Statistically aligned replicas of real intersections built in Unreal & Matlab RoadRunner. Geometry, materials, traffic flow, and sensor placement are all measured from the source — not guessed. The output is synthetic LiDAR that fools downstream detectors.
Training 3D object detectors entirely in simulation and benchmarking them on real roadside LiDAR. In several configurations the sim-trained detector now beats models trained on real data — the upside of perfect labels.
DINOSTAR trains object detectors on raw point clouds with no human labels — using statistically-modeled teachers built from clustering, background filtering, and bounding-box fitting. It matches human-annotated baselines.
LiGuard is a GUI-powered Python framework that lets researchers compose point-cloud pipelines without rewriting the framework. Published in JOSS, used by other labs, kept alive through three major versions.
pip install LiGuardFigures from the work — real intersections, the digital twins that copy them, and the perception models that train on the copies.
The projects I'd actually want to talk to you about. Research first, then tools, then the side projects that betray my game-dev origins.
A GUI-driven Python framework for LiDAR (and paired image) data. Researchers build dynamic pipelines by
toggling components in a config — no need to touch the framework's source. KITTI / OpenPCDet / SUSTechPoints
readers built in; custom processes drop into a single algo/ directory.
High-fidelity synthetic replicas of three real roadside LiDAR datasets (LUMPI, V2X-Real, TUMTraf-I). Built by aligning geometry, traffic, and sensor models to the source recordings. Lets detectors train on perfect labels and still generalize. Published as a dataset paper in IEEE OJITS 2026.
The Sim2Real LiDAR challenge at the 6th DriveX Workshop, IEEE/CVF CVPR 2026 (Denver). Submission built on the UrbanTwin pipeline. Co-author with Shaurya Agarwal.
A device-agnostic personal assistant. Android frontend (Jetpack Compose) talks to a FastAPI backend that orchestrates Gemini Flash, spawns workers (Thinking, Image-gen, Tool-Creator), and routes replies through a typed STM protocol with SDUI rendering.
A Windows tray app that turns my voice into keystrokes in any app. Live streaming transcription with LocalAgreement-2 commits — words appear while I'm still speaking, no flicker. GPU-accelerated, per-user settings, PyInstaller-packable.
A narrative VR game for Quest 3. Two protagonists, one mountain apartment, an AI-driven character with a bounded narrative and open-ended conversation. The fun side of "what happens when LLMs drive game NPCs."
A visual node-graph runtime for data pipelines — think ComfyUI for point clouds. React Flow canvas, streaming execution over WebSockets, parallel branch execution with proper error isolation, and a pluggable node registry.
More on GitHub — 39 public repos and counting.
UrbanTwin: Synthetic Roadside LiDAR Datasets
LiGuard: A Streamlined Open-Source Framework for Rapid & Interactive Lidar Research
Bridging Sim2Real Learning Gaps through High-Fidelity Digital Twins for LiDAR-based Perception Tasks
SpatioStar: A Framework for Modeling High-Fidelity Digital Twins for Sim2Real LiDAR-based ITS
DINOSTAR: Deep Iterative Neural Object Detector Self-Supervised Training for Roadside LiDAR Applications
Autonomous Drone Swarm Navigation in Complex Environments
DOI 2 citations
Translating LiDAR-based Object Detection from On-Road to Roadside: An Evaluatory Study
Self-Supervising Object Detection for Roadside LiDAR Applications
Using LiDAR for Real-Time Pedestrian Safety
University of Central Florida · Dept. of CECE · Urbanity Lab
Scaling Sim2Real methodologies for LiDAR perception. Two T-ITS-class manuscripts under review, one provisional patent, a CVPR challenge prize, and a fourth-version push of LiGuard.
UCF · CECE · FDOT-funded projects
Led PULSE (Pedestrian Upfront LiDAR-based Safety on Edge), worked on the Smart Work Zone project. 3D perception with LiDAR–camera fusion, sensor rigs (Ouster, Hesai, Flir), and the data infrastructure to feed them.
University of Central Florida · CGPA 3.71 / 4.0
Dissertation: From Failure to Fidelity: Enabling Scalable Sim2Real LiDAR Perception. ORCGS Doctoral Fellowship.
Pakistan Institute of Engineering & Applied Sciences (PIEAS) · CGPA 3.78 / 4.0
QS-ranked #1 university in Pakistan. Thesis on deep reinforcement learning for autonomous drone swarms; first-author IEEE paper at RAEE&CS.
Fiverr · Level 1 Seller
Led a smart-home build: lighting, thermostats, switches, curtains, a smart doorbell with dual audio/video. Raspberry Pi + Arduino Nano + cross-platform control software.
Personal · Google Play
Shipped Moneypad (personal finance), Blow (survival game), and 3D Banta Balls (online multiplayer take on the classic Pakistani street game). 1st in Regional Microsoft Imagine Cup, 2016.
PMAS Arid Agriculture University, Rawalpindi · CGPA 3.50 / 4.0
Senior design project: Augmented Reality Carrom Board with a custom conductive-thread glove controller. 1st place, 2018 Open House competition.
08 · Get in touch
I'm open to research scientist, applied scientist, and senior engineering roles in autonomous driving, robotics, and spatial AI. Postdoc collaborations welcome too.
Based in Orlando, Florida · open to relocation
