Trustworthy AI competition contribution

As a Technical Lead, drove product planning and built the core inference and monitoring stack for an LLM-based clickbait detection service, including 0/1 logit-based scoring, IG-based explanations, and short-/long-term drift detection with operational thresholds.

• Led service planning for a platform-style product that surfaces clickbait indices aggregated by publisher, journalist, and section, with an end-to-end user + operations flow.
• Built the end-to-end service pipeline from article ingestion to scoring, storage, aggregation, and monitoring-ready outputs.
• Designed and implemented the main inference engine by constraining outputs to {0,1}, extracting the final-layer 0/1 logits, applying a 2-class softmax, and persisting p0/p1 and margin scores for downstream use.
• Developed preprocessing and XAI modules, including JSON flattening/normalization for sliceable datasets and an Integrated Gradients (IG) module for token-level attribution.
• Fine-tuned the main inference model with DoRA, improving performance from [ accuracy [0.507] / precision [0.596] / recall [0.102] / F1 [0.174] ] to [ accuracy [0.986] / precision [0.985] / recall [0.988] / F1 [0.986] ].
• Owned impact analysis end-to-end, drafting the initial impact registry, prioritizing key impacts, and assigning roles/owners across the team.
• Proposed and executed long-term drift/bias detection, designing thresholding experiments and running an operations-style simulation to translate alerts into retraining/golden-set update decisions.
• Proposed and executed short-term drift/bias detection, designing and running experiments to calibrate practical alert thresholds and response actions.
• Produced and delivered the award-ceremony presentation, translating the system into a reliability narrative (prevention–detection–response) aligned with competition evaluation criteria.

Working as an AI model developer, specialized in enhancing adversarial robustness and operational stability by architecting a defense framework against Unicode-based obfuscation and a real-time short-term drift detection system.

• Built a news data collection and preprocessing pipeline, implementing text cleaning and NFKC normalization to ensure data consistency and integrity across the system.
• Proposed a detection methodology for Unicode-based obfuscation and conducted threshold-optimization experiments to identify and neutralize adversarial evasion attempts.
• Developed an operational simulation for short-term drift detection, using Negative Log-Likelihood (NLL) to monitor model stability against sudden linguistic shifts and emerging neologisms.
• Established an automated governance protocol with defined alerting and intervention workflows to safeguard system reliability during operational anomalies.

Working as a Data Scientist, supported interpretability analysis and validation for a clickbait detection system through label integrity checks, stress testing, and experimental reporting.

• Designed and refined XAI-based analysis experiments, clustering clickbait-related keywords at the human cognition level to produce interpretability-focused analysis and reporting.
• Contributed to the structuring and systematization of AI governance practices across data collection, preprocessing, modeling, and deployment, clarifying impact, risk, and response workflows.
• Conducted label integrity verification on a large-scale clickbait news dataset, quantitatively evaluating annotation reliability through inter-annotator agreement analysis.
• Designed and executed input-perturbation stress tests focused on punctuation variations, analyzing model vulnerabilities and operational risks, and documenting the results in validation reports.
• Led the creation and management of project-wide evidence artifacts, and producing presentation materials and a comprehensive project summary.

Working as a Data Scientist, conducted explainability-driven analysis and section-level bias evaluation for a clickbait detection system, and performed stress testing to identify and validate model failure modes under realistic conditions. Also led report authoring and contributed to evidence mapping and presentation material preparation.

• Designed and implemented Integrated Gradients–based XAI analysis code, and empirically validated model decision rationales through controlled experiments.
• Formulated section-level bias hypotheses based on label and subcategory distributions, and built a data proportion analysis pipeline from scratch to perform quantitative bias validation.
• Designed stress-test scenarios using disaster news datasets, modeling extreme cases where sensational lexicon is used in legitimate public-interest contexts, and verified model failure modes in which such cases are misclassified as clickbait.
• Led overall report and deliverable authoring, ensuring coherence and consistency across project outputs.
• Participated in systematic evidence mapping across experiments and analyses.
• Contributed to the development of presentation materials.

Working as Data Scientist, performed statistical bias and robustness analysis for a clickbait detection system, including dependency tests, stress testing, and fairness monitoring design.

• Defined and analyzed data bias issues in a clickbait article detection AI competition, statistically validating whether section-wise performance gaps stem from genuine section effects or structural confounding.
• Applied χ² tests and CMH conditional independence tests to disentangle section effects from processing-pattern confounders, and developed an effect-size – driven interpretation framework.
• Designed and conducted a section-swapping stress test to evaluate the impact of section–label dependency on model performance, confirming robustness to structural bias within the current data scope.
• Extended data and model analysis into an SLI/SLO-based fairness monitoring and risk management framework, proposing exposure-aware section-wise bias detection metrics and operational thresholds.