Research

Working title · Research in progress

My doctoral research investigates how automatic speech recognition systems might be adapted to clinical consultations involving older adults when relevant speech data and computing resources are limited. The proposed research brings together low-resource adaptation, on-device evaluation, and agent-guided experimentation.

The programme aims to examine recognition quality alongside practical deployment requirements, including latency and memory use on actual devices. A further research direction considers whether AI agents can support the selection and execution of adaptation experiments within defined data and computational budgets.

The current phase focuses on critically reviewing the literature, assessing the relevance of existing evidence, and refining research questions and evaluation protocols.

Accurate estimated time of arrival prediction is a practical challenge for food delivery platforms. This study uses DoorDash order data to examine feature engineering and compare Linear Regression, Random Forest, XGBoost, and CNN models for delivery duration prediction.

Feature engineering considers delivery distance, preparation-time bias, driver workload, and time-related variables. Order amount and item variety show weak correlations with delivery duration in the reported analysis.

Authors: Zeng Fan, Steve Teoh Chee Hooi, Saliyah Binti Kahar, Muhammad Tayyab, Israr Ahmad.

Conference: 2025 International Conference on Emerging Trends in Networks and Computer Communications (ETNCC).

Publication date: 5 August 2025. Pages: 701–705. Publisher: IEEE.

In the OLEAP App, I designed and explored a workflow linking transcript analysis, dynamic prompt construction, and meeting minutes generation. A large language model (LLM) identifies the meeting type and extracts topics from the transcript. This analysis is combined with a preset base prompt to construct meeting-specific instructions.

The transcript and the constructed prompt are used to generate meeting minutes. The research question is how content analysis can inform instructions that align the organization and information priorities with the discussion.

My contribution focuses on workflow design and incorporating meeting type and topic information into natural-language prompt construction. The description does not prescribe a particular prompt assembly mechanism or a fixed number of LLM calls.

This work is a method exploration based on my design account. Comparative evaluation against a fixed prompt and a meeting-type template baseline remains future work. No independently verified quality improvement is reported.

In an OLEAP corporate research and development project, I designed and developed a workflow that uses a large language model (LLM) to organize meeting content and generate a Markdown representation of a mind map. The method exploration focuses on information selection, hierarchical organization, and expression, with the aim of helping users understand the overall structure and key points of a meeting.

The research question is how to transform lengthy, linear meeting content into a browsable structured representation while preserving key information and its semantic relationships.

Further research could evaluate key content coverage, the coherence of hierarchical organization, information redundancy, and content faithfulness. Such evaluation could examine whether outputs omit important information, introduce content unsupported by the source, or alter the original meaning.

These are directions for future empirical validation, rather than established experimental conclusions.

In an OLEAP corporate research and development project, I designed and implemented a mobile recording workflow for framed audio received from a wearable device over Bluetooth. The prototype connects recording events, received-block checks, threshold-triggered buffering, append writes, and normal-stop handling to organize local recording data and metadata. It provides an input foundation for recording management and subsequent transcription, rather than ASR model training or acoustic algorithm development.

The pipeline distinguishes an audio frame, a received block that may contain several frames, and a write batch combining multiple blocks. The threshold counts received blocks rather than a fixed number of frames or a fixed time window. Block-length checks test format compatibility, not whether frames are missing, duplicated, or corrupt.

At the threshold, buffered blocks are combined in arrival order and submitted for append writing. On normal stop, the remaining partial batch is submitted and recording metadata is saved. Duration is estimated from the accumulated frame count and an assumed frame duration; it does not establish hardware timestamp synchronization or forced alignment.

Long recordings are the target use case, not a validated result. Maximum recording duration, peak memory use, frame loss, and performance have not been measured. Batch buffering alone does not establish bounded memory across the workflow: final conversion reads the full file and creates additional copies.

The current prototype does not establish serialized write completion or draining of in-flight writes. It lacks an independently bounded producer-consumer queue and backpressure; frame-sequence recovery, retransmission, write retry, and crash recovery are not established capabilities. Normal-stop handling does not guarantee loss-free recording or crash-safe persistence.

Opus decoding remains incomplete and is not claimed as an implemented codec contribution. Future evaluation could compare block-count, byte-count, and time-based triggers against per-block writes, measuring application-level write calls, data residence time, memory use, write ordering, and data integrity. No algorithmic novelty or measured performance improvement is claimed.

Conducted user-centered research and product design to address accessibility barriers faced by visually impaired massage clinic owners in China. Investigated how conventional Point-of-Sale (POS) systems fail to accommodate screen-reader interactions, and explored accessible mobile app interface design, non-visual navigation, transaction workflows, and business reporting to support independent business management.

Recognition: Bronze Award, 6th China International “Internet+” College Students’ Innovation & Entrepreneurship Competition — Guangdong Provincial Competition, Youth Red Dream Journey Track (2020).