Projects

Problem

A query can run successfully and still select the wrong columns or join the wrong tables. I compared the original and adapted models on the same questions, scoring executability and result correctness separately.

Methods

• I fine-tuned the model on 800 examples using QLoRA on one 16 GB V100. The reference SQL syntax trees supplied labels for tables, columns and join edges. A full-vocabulary KL penalty anchored training to the frozen base model.
• During generation, a token-level grammar and alias-scope check blocks invalid continuations. When static checks flag column ownership or a join path, the model gets one new attempt using the question, full schema and diagnostic feedback.

Results

On the same 100 BIRD development questions, the base model returned 36 correct results and 71 executable queries. Fine-tuning with constrained decoding brought these counts to 44 and 75. One repair attempt made two more queries executable, with 44 correct results overall.

Problem

Functional data such as multilead ECGs are high-dimensional and strongly correlated. How can shared and function-specific risk components be learned under right-censored survival outcomes while retaining traceable interpretations?

Methods

• Extend multiscale shared representation learning to right-censored survival outcomes. Use the score and information geometry of Cox partial likelihood to group leads, extract shared and specific risk components, and learn subsequent risk directions through residualization.
• Implement Fusion and Directional grouping using agreement in lead-level risk contributions and standardized risk directions, respectively. Recursive transformations reconstruct predictions as lead-by-time loadings to explain contributions across leads and time.
• Use CODE-15 twelve-lead ECG data in multiple split seeds, five-fold validation and nested training sizes to examine representation complexity and risk discrimination.

Results

The study includes 154,744 twelve-lead ECG records and 7,888 deaths. Across five-fold validation with 10 split seeds, full-size Directional models reduce the mean component count from 72 to 36.48, with a mean C-index of 0.74598. Supplementary Fusion results are 25.02 components and a C-index of 0.74548.

C-index uses the Uno estimator. These summaries describe existing repeated-split results and do not establish improved predictive performance or statistical equivalence. Lead associations describe fitted model structure, not physiological connections or causal relationships.

Problem

How can follow-up information support prognosis modeling and comparisons of sequence models and survival methods within a leakage-controlled evaluation pipeline?

Methods

• Integrate diagnosis, treatment and progression events for 5,368 MSK-CHORD breast cancer patients using SQLite. Construct landmark samples with patient-level splits and training-side preprocessing to predict subsequent risk at different follow-up times.
• Represent variable-length event sequences with categorical embeddings and numerical features. A Transformer encoder and masked pooling learn patient representations with a numerically stable Cox loss for right-censored outcomes, using only events before the prediction time.
• Examine how predictions respond to changes in disease-progression variables through model sensitivity analysis.

Results

Develop longitudinal risk modeling and interpretation methods, combining SHAP and progression-event perturbations to examine model sensitivity while distinguishing predictive associations from causal effects.

There is no formal external or prospective validation. Event modifications assess model sensitivity, not identified causal treatment effects.

Research question

Among patients with non-small cell lung cancer (NSCLC) who received neoadjuvant immunochemotherapy followed by curative surgery, investigate the predictive value of primary-tumor and lymph-node features from pre- and post-treatment chest CT for patient-level postoperative pathological nodal status (ypN).

Work to date

• Combine SLIC local partitions and multilevel features from pretrained ResNet18 with PCA and GMM fitted within training folds to build a 102-dimensional patient representation of local heterogeneity and lymph-node sets.
• Integrate 50 spatial and longitudinal features to model tumor and lymph-node phenotypes and pre/post-treatment changes. Evaluate discrimination, calibration and generalization in 827 patients using patient-level ten-fold and three-center leave-one-center-out validation.
• Supporting data audit: of 14,775 lymph-node connected components, 3,391 (about 23%) did not reach feature extraction because of resampling loss or filtering rules, informing improvements to small-lesion representation.

Current stage

The study is ongoing, focusing on reliable baselines, identifying information loss, and improving representations. Current candidates do not yet support stable adoption; no final model has been selected.

Problem

Simulate Hong Kong unsecured installment lending and turn ongoing customer, application, disbursement, and repayment events into consistently defined, traceable daily business metrics.

Methods

• Define approval rate, disbursement, loan balance and delinquency metrics in Python and PostgreSQL, distinguishing flows from stocks. Handle duplicate, late and corrected records through event versions, incremental computation and reconciliation.
• Schedule daily processing and correction-window rebuilds locally, publishing only aggregate metrics to the cloud for a five-page Streamlit dashboard.
• Add data-quality and reconciliation checks; verify consistent reruns, correct propagation of revised records into metrics, and transaction rollback on failure.

Results

The verified project snapshot covers 248 days, 4,665 disbursed loans, and HK$94.7M in principal, with 19 passing tests. The dashboard covers the executive overview, credit funnel, portfolio and collections, origination vintages, and channel performance and data quality.

All figures refer to synthetic data and a project acceptance snapshot, not real customers, actual lending volume, or the dashboard’s current live values.

Problem

Who should receive an SMS when the budget covers only 30% of customers? Using 200,039 labeled customers, distinguish those likely to buy anyway from those whose purchase probability may increase because of the message.

My work and model design

• Reconstruct receipts from transaction lines and build 53 customer features covering RFM, recent spending trends and points behavior. Fix seed 42 and training, validation and evaluation partitions; compare random sending, RFM, response prediction and uplift policies.
• Design C1, a regularized DR-learner: combine purchase-probability models with treatment-assignment information to form uplift targets, then learn customer uplift using complexity-constrained gradient-boosted trees. The final model uses at most 7 leaves and at least 1,000 customers per leaf to limit fitting to noisy targets.
• C1 achieved the highest validation AIPW value among eight candidate policies and was frozen as the final policy. Rank positive uplift within the 30% budget, examine list overlap and customer profiles, then evaluate offline using 40,083 held-out customers.

Results

On narrow screens, scroll horizontally to view all columns.

Intervals use Bonferroni correction for two comparisons (97.5% per comparison; family α=0.05). AIPW provides nominal positive evidence, while both IPW intervals include zero. Estimates assume random assignment with probability 0.5. This round reused previously examined historical evaluation data without adjustment for exploration across rounds; it is neither independent confirmation nor measured online impact.

Tree depth, direct budget optimization and feature-count comparisons remain supplementary research. Evaluation results do not replace the frozen C1 policy, and non-significance is not treated as equivalence.