Work

Work you can interrogate.

These are completed products and personal research engineering projects. Each one separates what was reproduced, what changed, and what evidence remains.

Academic citation verification across heterogeneous sources

A reference can be real and still be wrong for the sentence using it. The system needed to distinguish existence, metadata accuracy and evidential relevance.

The problem

Citations arrive as incomplete prose, DOI strings, mixed formats or model-generated text. Individual scholarly APIs have different coverage and identifiers.

The system

Parse and normalize the input, search five scholarly sources, reconcile candidates, compare fields and present confidence without hiding uncertainty.

The evidence

Users see source links, title and author differences, publication metadata and an explicit explanation of what was or was not verified.

Multi-source retrieval pipeline
Field-level mismatch detection
Rate-limit and partial-result handling
Production web integration

Personal research archive

From paper idea to measured result.

Summaries are grounded in completed code, experiment logs and delivery packages. They are anonymized; private data, weights and unpublished method details stay private.

PINN

Scientific ML / Physics

Ocean acoustic pressure magnitude predicted by a physics-informed neural network across range and depth — Predicted pressure magnitude from the KAN model with a hard boundary constraint.

Physics-informed ocean acoustic field prediction

Reproduced a published pressure-field method, introduced a KAN backbone with hard boundary constraints, and aligned evaluation against Kraken ground truth.

Built: Rebuilt the baseline, stabilized PDE training, ran more than 12 ablations, and packaged repeatable training and evaluation.
Evidence: Matched baseline accuracy in the interpolation region with 35k parameters instead of 130k, while reporting the remaining extrapolation gap.

VIS

Computer vision / Detection

Ablation comparison for aerial small-object detection showing inputs, ground truth and four response heatmaps — Input, ground truth and response heatmaps from four detector configurations.

Small-object detection in aerial scenes

Implemented a self-relation detector for the SIMD aerial dataset, focusing on small targets and rare-class imbalance.

Built: Unified the training path, added gated residual fusion and rare-class sampling, and delivered weights plus one-command inference.
Evidence: Best mAP50 increased from 0.7269 to 0.7584 under the same evaluation setup.

DAG

Edge computing / Optimization

Six-way distributed DNN placement ablation showing the full method with the lowest latency and 63.7 percent improvement — Latency ablation for resource, communication and merge-aware modules.

Model placement for distributed DNN inference

Reproduced a collaborative-edge placement algorithm and addressed local branch optima and underspecified merge-node waiting.

Built: Implemented multiple baselines, DAG critical-path evaluation, six ablations, and seeded batch experiments.
Evidence: In the explicitly labelled calibrated scenario, the full method reduced mean latency by 63.7% against the reproduced base.

FUS

Remote sensing / Image fusion

Multispectral and panchromatic image fusion

Built a cross-dataset training, inference and metric pipeline around FusionMamba and PanFormer.

Built: Covered WV3 and GF2 at reduced and full resolution, isolating the contribution of frequency and attention modules.
Evidence: Eight complete ablations produced 320 replayable outputs with matching metrics and generation logs.

MMF

Multimodal ML / Emotion

Macro-F1 comparison of text-only, dynamic-weight fusion and text-anchored fusion on expanded 15-speaker and 24-speaker datasets — Text baseline versus two fusion strategies on the expanded datasets.

Text, speech and expression fusion

Built a text-anchored multimodal emotion experiment focused on speaker-independent generalization and missing-modality robustness.

Built: Implemented LOSO evaluation, five seeds, unimodal and fusion baselines, weight ablations, and report generation.
Evidence: Macro-F1 on the expanded experiment rose from a 0.471 text baseline to 0.520 with fusion.

NLP

NLP / Information extraction

Nebula-style knowledge graph of anonymized filing entities and relations; aggregate metrics remain visible.

Joint entity and relation extraction from Chinese filings

Adapted an autoregressive text-to-graph method to long Chinese filings and tested where type-constrained candidate pruning actually helps.

Built: Converted 651 annotations, adapted the Chinese encoder, implemented constrained decoding, and retained both positive and negative ablations.
Evidence: The converged model reached 84.97% entity F1 and 55.53% relation F1; unsuccessful ideas remain documented.

Have a paper, algorithm idea or experiment that is stuck? Start with one small question we can verify.

Explore research support