Research and Technology Forum cum Get-together Event 2026

Edge Intelligence: Multi-Modality Sensing to Revolutionize Chronic Diseases Management

Rapid aging has led to an increase in the number of chronic disease patients, rising social medical costs, and a decline in the quality of life of patients. Under today's hospital-centered service model, patients are already seriously ill when diagnosed, and their subsequent prognosis is quite uncertain. Our research attempts to transform the service model into a lifecycle management model, using IoT sensors and AI algorithm design for early diagnosis of diseases, and will also provide continuous health assessment and rehabilitation guidance to reduce the cost of medical services.

I will share our recent work on chronic disease assessment and intervention supported by intelligent sensing (multiple sensing modality will be leveraged) and AI algorithm design. I will use COPD assessment and training (leveraging audio, IMU, as well as depth camera), as well as dry eye disease screening and management (leveraging mmWave), as example for the sharing.

Securing the Agent Economy: Building the Trust Infrastructure for Web 4.0

As the internet evolves from Web 3.0's "Read-Write-Own" to the Web 4.0 paradigm of "Read-Write-Own-Earn-Transact," AI agents are becoming autonomous economic participants. However, while agent-managed on-chain assets are growing 10x year-over-year, the safety infrastructure remains almost nonexistent. This talk explores the emerging "Agent Attack Surface," where language models can be persuaded into irreversible financial transactions without a human in the loop. We present a preliminary defense architecture—from encrypted credential storage to on-chain constitutional anchoring—to ensure that only compliant, verifiable agents can participate in the future global economy.

Formal Approaches for Software Migration and Evolution

Software now plays a central role in many aspects of human society. Current software development practices involve significant developer effort throughout the software life cycle, including the development of new software, continuous improvement of code structures, maintenance of legacy software, and integration of existing software into more contexts. In this talk, I will discuss three automated formal techniques designed to streamline these processes: dynamic program inference, static semantic analysis, and syntactic adaptation. These approaches offer diverse strategies for addressing the challenges inherent in software migration and evolution. Our goal is to reduce manual effort and improve software quality across various contexts and environments.

Efficient Spatial and Embodied Intelligence under Resource Constraints

As AI systems increasingly interact with the physical world, efficiency becomes a system-level challenge spanning computation, memory, and actuation. This talk presents our recent work on rethinking efficiency for spatial and embodied intelligence under limited resources. We introduce an out-of-core training framework that enables billion-primitive 3D Gaussian Splatting on a single consumer GPU, and revisit efficiency in Vision-Language-Action models for robotic manipulation, showing that conventional inference-centric metrics fail to capture real-world power and execution costs. By proposing embodied efficiency metrics that reflect physical behavior, these works highlight the need for algorithm–system–hardware co-design and suggest that future AI systems should be evaluated not only by how fast they compute, but by how efficiently they interact with the physical world.

Private Information Retrieval: From Theory to Practice

Information retrieval is a cornerstone of the modern Internet — from web search engines like Google to multimedia platforms like YouTube and TikTok, and biomedical applications such as fingerprint and facial recognition. Yet traditional search architectures expose users' search intent to service providers, raising serious privacy concerns. In 1995, Chor et al. introduced Private Information Retrieval (PIR), a cryptographic primitive that enables a client to retrieve records from a server without revealing which record was accessed. While the decades since have produced rich theoretical developments, building practical PIR systems that scale to real-world applications remains a profound challenge.

In this talk, we present our series of works toward making PIR practical at scale, based on research papers featured in top-tier cryptography and system security conferences (Eurocrypt 2023, 2024, 2026, IEEE S&P 2024, 2026). Building on the Preprocessing PIR model (Corrigan-Gibbs and Kogan, Eurocrypt 2022), we develop a line of new constructions incorporating several technical innovations that have shaped the research frontier. As a centerpiece, we present Piano, the first practical single-server preprocessing PIR scheme with amortized sublinear computation and communication per query. Piano achieves sub-20ms query latency on databases with billions of records — a two-to-three order-of-magnitude improvement over prior solutions. Extending this foundation, we further present Pacmann (ICLR 2025), a private semantic search system built on Piano that delivers over 0.9 recall@10 on a vector database of 100 million entries within 2 seconds.

Revisiting RDMA Reliability for Lossy Fabrics

RDMA is widely used in modern datacenters but typically depends on lossless networks with Priority Flow Control (PFC), which limits scalability and increases operational complexity. In this talk, I present DCP, a new transport architecture that enables efficient and scalable RDMA over lossy Ethernet fabrics. DCP is independent of PFC, compatible with packet-level load balancing, free from retransmission timeouts, and designed for hardware offloading. By co-designing switches and RNICs, DCP introduces header-only retransmissions and lightweight packet tracking. A P4 switch and FPGA prototype demonstrate up to 2× performance improvements over existing RDMA solutions while significantly enhancing scalability.

PhD (CSE) Graduate (Class of 2026)
​ACM SIGCOMM '25 Best Student Paper Award (Honorable Mention Recipient)

Contactless Arterial Blood Pressure Waveform Monitoring with mmWave Radar

Arterial blood pressure waveforms (ABPW) provide richer cardiovascular insights than discrete SBP/DBP, but existing monitoring methods are invasive or require continuous skin contact. We propose WaveBP, the first contactless ABPW monitoring system utilizing a commercial mmWave radar, driven by the understanding that cardiac information serves as an implicit bridge between mmWave signals and ABPW based on a hemodynamics analysis model. WaveBP reconstructs beat-to-beat waveforms with mmFormer, a hybrid Transformer with spatially informed shortcuts that supports flexible personalization. To improve robustness against mmWave instability, we introduce a beamforming-based multi-view augmentation, and further boost accuracy via cross-modality knowledge transfer from ECG/PPG during training with no extra deployment overhead. On 43 subjects under LOSO evaluation, WaveBP achieves 0.903 waveform correlation and (-0.14±7.48) mmHg point-level error, meeting practical accuracy requirements and enabling downstream detailed cardiac estimations such as relative cardiac output and abnormality detection in a contactless manner.

ACM IMWUT Distinguished Paper Award Recipient (UbiComp/ISWC)

From Human-centric to Model-centric: Reimagining Agentic Skills in the Age of LLMs

The rapid evolution of large language models and the surging volume of agentic API calls, as exemplified by platforms like Anthropic and OpenRouter, signal a fundamental shift in software development. However, current agentic systems often rely on human-centric skill designs that lead to significant security vulnerabilities such as poisoned tool invocations and inconsistent performance across various models. In this talk, I will introduce our recent research on Model-centric Skill Engineering. We propose that agentic skills should be optimized specifically for LLM reasoning patterns instead of human readability. I will present our technical framework for skill optimization and demonstrate through experimental results how this approach enhances both safety and task success rates in real-world deployment.

PhD (CSE) Graduate (Class of 2025)
Ant InTech Scholarship Recipient

PathAdvisor 3.0: Real-Time Infrastructure-less Indoor Localization on Edge Devices

Indoor localization remains a fundamental challenge in modern smart environments. While the Global Positioning System (GPS) work effectively outdoors, their performance significantly degrades indoors due to signal blockage. Many existing indoor positioning solutions rely on Wi-Fi or Bluetooth infrastructure, which requires costly deployment, high energy consumption, and frequent maintenance in large or complex buildings.

This presentation introduces an infrastructure-free indoor localization system. The system relies solely on data from the inertial sensors already available in mobile devices and combines multiple sensing techniques to estimate user movement and position. Designed for real-time operation on edge devices, the approach enables practical indoor positioning without requiring additional hardware.

2025-26 CSE Best Final Year Project (FYP) Awardee