Selected work

Representative Papers

A Cascaded Unsupervised-Supervised NLP Pipeline for Detecting Accusatory Language in Public Procurement

Aug 12, 2026

This study addresses the underutilization of stakeholder comments and government open data in public procurement, which hinders timely detection of procedural violations. To overcome this challenge, the authors propose a lightweight, domain-adapted cascaded unsupervised–supervised NLP framework. The approach first employs domain-finetuned Word2Vec embeddings combined with Gaussian Mixture Model (GMM) clustering to uncover latent anomalous patterns; it then applies a Random Forest classifier to identify comments with accusatory intent. Despite severe class imbalance, the method achieves high precision and recall without requiring extensive computational resources, enabling effective identification of procurement risks and significantly enhancing regulatory transparency.

0 citationsRead paper

The Plausibility Trap: Using Probabilistic Engines for Deterministic Tasks

Jan 21, 2026

This study addresses the prevalent misuse of generative AI in tasks that could be efficiently solved by deterministic methods, leading to substantial resource waste and the risk of “algorithmic flattery.” Introducing the novel concept of the “reasonableness trap,” the work quantifies this misapplication through microbenchmarks and case studies in domains such as OCR and fact-checking, revealing approximately 6.5× higher latency overhead and reduced reliability. To mitigate this issue, the authors propose a tool selection engineering framework alongside a deterministic–probabilistic decision matrix, offering developers systematic guidance for appropriate technology choice. The paper further advocates integrating digital literacy education with the critical competency of discerning when not to deploy AI, emphasizing judicious tool adoption over default reliance on generative models.

0 citationsRead paper

Recent publications

Latest Papers

A Cascaded Unsupervised-Supervised NLP Pipeline for Detecting Accusatory Language in Public Procurement

Aug 12, 2026

This study addresses the underutilization of stakeholder comments and government open data in public procurement, which hinders timely detection of procedural violations. To overcome this challenge, the authors propose a lightweight, domain-adapted cascaded unsupervised–supervised NLP framework. The approach first employs domain-finetuned Word2Vec embeddings combined with Gaussian Mixture Model (GMM) clustering to uncover latent anomalous patterns; it then applies a Random Forest classifier to identify comments with accusatory intent. Despite severe class imbalance, the method achieves high precision and recall without requiring extensive computational resources, enabling effective identification of procurement risks and significantly enhancing regulatory transparency.

0 citationsRead paper

The Plausibility Trap: Using Probabilistic Engines for Deterministic Tasks

Jan 21, 2026

This study addresses the prevalent misuse of generative AI in tasks that could be efficiently solved by deterministic methods, leading to substantial resource waste and the risk of “algorithmic flattery.” Introducing the novel concept of the “reasonableness trap,” the work quantifies this misapplication through microbenchmarks and case studies in domains such as OCR and fact-checking, revealing approximately 6.5× higher latency overhead and reduced reliability. To mitigate this issue, the authors propose a tool selection engineering framework alongside a deterministic–probabilistic decision matrix, offering developers systematic guidance for appropriate technology choice. The paper further advocates integrating digital literacy education with the critical competency of discerning when not to deploy AI, emphasizing judicious tool adoption over default reliance on generative models.

0 citationsRead paper