Your system preference limits motion.
RESEARCH REVIEWER ENTRY
I work on frontend systems and data quality, and study reliable multimodal evaluation. This page connects my first-author submissions, public engineering work and proposed doctoral questions.
Status compiled 2 November 2026 from the supplied manuscript information and portfolio records.
When labels and compute are limited, can multimodal improvements survive changes in validation, modality quality and decision thresholds? For proposed doctoral work, I want to investigate this question in reliable emotion recognition.
The current fake-news study motivates the evaluation question. Moving to emotion recognition requires new datasets, tasks and controlled experiments; it is not an already validated transfer result.
Read the existing evidence first, then the proposed first study. The climate project is an interdisciplinary application; the proposed doctoral question remains reliable multimodal evaluation.
Existing evidence ↗First experiment ↗First author · Submitted, not accepted
The revised study reanalyzes 344 archived training runs and examines fusion gains, operating points and archive artifacts. The earlier submission was rejected; the revision was submitted to Neurocomputing.
Public analysis code and saved predictions are available. On 2 Nov, 216 few-shot operating-point rows were recomputed and matched the reference. Training and original split reconstruction remain outside this check.
First author · Submitted, not accepted
Hybrid AI Framework for Multi-Hazard Climate Risk Governance in Renewable Energy-Water Nexus Systems. The supplied abstract combines TFT and GNN-CRP for 30 Chinese provinces over 2005–2024.
Available: the abstract-based overview and a conceptual method map. This page does not present a public benchmark package or add unprovided performance figures.
Shared state, upload lifecycle and versioned release records. Local pilot; real-device acceptance remains open.
An AI-assisted React + Ant Design learning example. Eight pure model tests rerun on 1 Nov 2026.
A custom schema learning example. Ten core tests rerun on 1 Nov; browser input retention is checked separately.
These checks demonstrate bounded engineering behavior. They do not establish a new research result or production deployment capacity.
Count training, model-selection and calibration labels together; compare under matched budgets.
Separate modality quality, disagreement and calibrated confidence; examine selective prediction.
Fix replay and update budgets, and evaluate both new-task performance and forgetting.
These are proposed studies awaiting experiments. Engineering experience supports data checks, repeatable experiment interfaces and traceable versions; controlled research must establish the method claims.
First milestone: compare unimodal and simple fusion baselines under shared train/selection/calibration budgets. Freeze data splits and a primary endpoint before observing test results.
Read the proposed protocol ↗