CV

Download PDF Resume

Education
University of Illinois Urbana-Champaign
PhD Candidate in Computer Science
Advised by Prof. Klara Nahrstedt. Previously worked with Prof. Deepak Vasisht and Prof. Matthew Caesar.
Research: Spatial Intelligence, On-Device ML, Wireless and Mobile Systems.
University of Illinois Urbana-Champaign
Master of Computer Science
University of Illinois Urbana-Champaign
B.S. in Statistics & Computer Science, Highest Distinction
Experience
Google — Android Location Team
Student Researcher (Intern)  ·  Mountain View, CA
Supervised by Dr. Roy Want. Android Location Team — the team behind indoor positioning in Google Maps and Android Location Services.
  • Advanced the integration of Wi-Fi Round Trip Time (RTT) indoor localization into Android's Fused Location Provider (FLP) and scaled evaluation across numerous Google offices in Mountain View.
  • Achieved sub-second localization latency with 1 m 90th-percentile error on Google Wi-Fi access points and 2 m on enterprise-grade HPE Aruba, Cisco, and Arista wireless access points.
  • Developed and evaluated tightly coupled particle-filter and Bayesian-grid methods for fusing Wi-Fi RTT measurements with motion sensors across varied motion and ranging conditions.
  • Designed and implemented APSelector, a geometry-aware AP selection algorithm that prioritizes access points providing strong geometric constraints to mitigate NLOS and positive ranging biases.
  • Incorporated IMDF walkable-path constraints, when available, as probabilistic priors to reject infeasible trajectories and further improve localization accuracy.
Google — Android Location Team
Student Researcher (Intern)  ·  Mountain View, CA
Supervised by Dr. Roy Want. Android Location Team — the team behind indoor positioning in Google Maps and Android Location Services.
  • Designed and developed an advanced indoor localization prototype combining Wi-Fi Round Trip Time (RTT) and motion sensor fusion on Pixel phones, achieving substantially lower latency and higher accuracy than Google's production Fused Location Provider (FLP) under challenging indoor environments (DFS channels, multipath-heavy areas).
  • Implemented a passive Wi-Fi listening framework that reuses cached scan results from FLP to enable continuous RTT-based ranging without active scans, reducing localization update latency by 57% and eliminating blackout periods.
  • Built a high-precision pedestrian dead reckoning (PDR) module leveraging Android's step detector and orientation sensors, and fused it with Wi-Fi RTT measurements through a multi-state Kalman filter, achieving 0.8 m indoor localization accuracy at 90% CDF using existing commercial Wi-Fi infrastructure.
  • Contributed to ongoing efforts to integrate Wi-Fi RTT capabilities into Google's Fused Location Provider framework for next-generation indoor localization.
Publications
Enguang Fan, Binh Minh Tran, Fulin Wang, Jiaxi Li, Klara Nahrstedt
Accepted to the NeurIPS 2026 PhysWorldAI Workshop.
RTT-MOTION: Scan-Aware Real-Time Indoor Localization with Commodity Wi-Fi RTT
Enguang Fan et al.
Under submission to ACM MobiCom 2027.
Enguang Fan*, Binh Minh Tran*, Klara Nahrstedt
32nd ACM Annual International Conference on Mobile Computing and Networking (MobiCom '26), short paper. * Equal contribution.
Enguang Fan, Yifan Chen, Zihan Shan, Matthew Caesar, Jae Kim
Accepted to the 2026 IEEE Global Communications Conference (GLOBECOM).
Enguang Fan, Anfeng Peng, Matthew Caesar, Jae H. Kim, Josh Eckhardt, Greg Kimberly, Denis Osipychev
IEEE Military Communications Conference (MILCOM), Boston, USA.
Emerson Sie*, Enguang Fan*, Federico Cifuentes-Urtubey, Deepak Vasisht
arXiv preprint. * Equal contribution.
Enguang Fan*, Emerson Sie*, Federico Cifuentes-Urtubey, Deepak Vasisht
Poster at the 31st ACM Annual International Conference on Mobile Computing and Networking (MobiCom), Hong Kong, China. Best Poster Runner-Up. * Equal contribution.
Enguang Fan
arXiv preprint arXiv:2602.01673.
Research Experience
On-Device Multimodal GUI Agent for Android
  • Built a multimodal Android GUI agent that executes natural-language, cross-app tasks with a vision-language model (VLM) run fully on-device (Qwen2.5-VL-3B-Instruct, W4A8, on a Snapdragon 8 Gen 3 phone) or in the cloud (Gemini Flash), behind one shared agent, tool, and recovery interface.
  • Designed an observe-reason-act loop on Android AccessibilityService that turns the UI tree into indexed observations, grounds JSON-schema-constrained actions (tap, swipe, type, launch, navigate), falls back to screenshot perception for inaccessible UI, and re-validates actions against the live UI to recover from stale targets.
  • Benchmarked on-device execution on real cross-app workflows, measuring median 12.1 tokens/s decode, 2.9 s vision encoding, and 10.8 s end-to-end action latency (vs. 4.1 s in the cloud) at 3.6 GB peak memory; evaluated Hexagon NPU offload via Qualcomm QAIRT/QNN and kept the ARM CPU as the stable full-model path.
Learning Latent RF Landmarks from Unlabeled Mobile Trajectories
  • Developed a self-supervised Radio Frequency (RF) place-recognition framework that learns discriminative spatial embeddings from sparse Wi-Fi RTT/RSS trajectory segments, enabling mobile devices to discover repeatable latent landmarks without labeled locations or known access-point coordinates.
  • Introduced RF-based loop closure for crowdsourced mobile SLAM, matching independently collected trajectory segments and incorporating high-confidence RF correspondences as constraints in pose-graph optimization to correct accumulated pedestrian dead-reckoning drift.
  • Designed an end-to-end mapping pipeline that jointly reconstructs mobile trajectories and reusable RF landmarks from unlabeled walks, with robustness to missing AP observations, multipath, trajectory direction, and cross-device measurement variation.
  • Built and released a synchronized Wi-Fi RTT/RSS/IMU trajectory dataset and reproducible SLAM benchmark with ground-truth poses for evaluating RF place recognition, loop-closure detection, and crowdsourced spatial mapping.
Crowdsourcing Ubiquitous Indoor Localization with Wi-Fi Ranging
Supervised by Prof. Deepak Vasisht, UIUC
  • Designed and implemented PeepLoc, a wireless indoor localization system fusing non-cooperative Wi-Fi ranging with pedestrian dead reckoning (PDR) without infrastructure support or PHY-layer access, achieving 56% lower mean error than Android's Fused Location Provider (3.41 m vs. 7.71 m) across four buildings.
  • Proposed a probabilistic backend that geolocates access points by fusing one-way time-of-flight (ToF) estimates with PDR trajectories in a joint non-linear least squares problem under clock-offset uncertainty.
  • Developed a per-AP adaptive ranging model correcting non-line-of-sight (NLOS) multipath distortion and hardware-induced RTT slope deviations, improving end-to-end localization accuracy indoors.
Honors & Awards
* **MobiCom 2025 Best Poster Runner-Up** * UIUC Fall 2022 Teachers (TA) Ranked as Excellent by Their Students — [Link](https://citl.illinois.edu/docs/default-source/teachers-ranked-as-excellent/tre-2022-fall.pdf)
Skills
* **Languages**: Python, C, C++, Java, MATLAB * **Machine Learning, Deep Learning & Computer Vision**: PyTorch, TensorFlow, Transformers, LLMs/VLMs, on-device inference, quantization, LSTMs, Reinforcement Learning, OpenCV * **Robotics & State Estimation**: SLAM, visual-inertial odometry, factor graphs, Kalman/particle filters, sensor fusion, PDR, ROS/ROS2 * **Wireless & Mobile Systems**: Wi-Fi RTT, Android, Linux, CMake