Academic project for the Ministry of Justice’s Kamlangjai program. The workflow supports aggregate-only analysis and hypothesis generation; it does not make individual classification or decisions.

• Data: linked 12 corrections datamart tables with probation data in DuckDB into person-level Parquet tables covering 2M+ people, with a data contract that tags features as usable, proxy, or protected.
• Analysis: compared occupational-grant recipients with non-recipients using direct standardization by sex, age band, and prisoner class, then applied Benjamini–Hochberg correction across tested factors.
• Rule-consistency check: benchmarked multinomial logistic, ordinal, and gradient-boosting models against a majority baseline for prisoner class. The leading features were inputs to the official rule, so the models are reported as a rule-consistency check, alongside a fairness audit and model card.

A ranking study for MITRE ATT&CK techniques associated with CVE roles. The public main branch contains the earlier W3 foundation; reported research results are pinned to the work/w8-ablations commit.

• Reported sealed-test evaluation: hybrid nDCG@5 0.5581 across 158 queries, compared with text-only LambdaRank 0.5317, MiniLM 0.2178, and BM25 0.1451.
• Interpretation: 5×5-fold ablations found that the gain over text-only ranking came from platform (CPE) features. A randomized graph scored comparably to the real graph, so the result should not be interpreted as evidence that graph structure supplied the gain.

A coauthored study of 20,919 undergraduate student records from Sakon Nakhon Rajabhat University, Thailand (2014–2024). The research develops staged early-warning predictions after Terms 1, 2, and 3+ from student-level academic data.

• Context: raw course-level records were cleaned, aggregated to credit-weighted term GPA, joined with student and academic tables, and pivoted across up to eight terms.
• Approach: feature engineering included GPA trend, variation, failure, threshold, and interaction features. Logistic regression, random forest, and XGBoost were assessed with an 80/20 student-level split and cross-validation on training data.
• Reported evaluation: using an 80/20 student-level split, XGBoost led on ROC-AUC and F1 at each checkpoint. The supplied resume reports ROC-AUC 0.866 after Term 1 and 0.992 after Term 3. These predictive results do not establish an intervention effect.

A local-first Windows SQL workspace for exploring files without treating a local dataset as a remote service. It is designed around a reviewable path from opening a file through reading results and exporting a new CSV.

• Workflow: open local CSV, TSV, Parquet, JSON, or JSONL files → write a SELECT or CTE query → inspect the live preview → execute and page through results → export a new CSV while preserving original files.
• Approach: Electron, React, TypeScript, Monaco, and native DuckDB support scoped completion, a visual ER/JOIN builder, profiling, reviewed Filter-to-SQL, and saved query and data tabs.
• Validation: its 21 September 2026 README reports 264 native tests and separate 9/9 packaged workflow checks. Remote connectors, DML/DDL, and data upload are outside its current scope.