Niles Liu

Generative modeling on manifolds

Learning & Adaptive Systems Group · Dr. Ya-Ping Hsieh

The challenge was imbalance. Training distributions are long-tailed across protein families, so diffusion samplers drift toward the modes they have seen most. I proposed a trajectory-consistency objective, called adjoint matching, that aligns forward and reverse dynamics to encourage exploration, then derived a manifold-aware variant for SE(3)-constrained spaces. In controlled tests it recovered minority modes that standard samplers collapse without trading away accuracy.

Calibrated UQ for clinical machine learning

Institute for Medical Engineering & Science · Dr. Li-Wei H. Lehman

In the ICU, a false arrhythmia alarm is a calibration problem before it is an accuracy one. The cost of crying wolf and the cost of a missed event are nothing alike. I used deep-ensemble fully-convolutional networks and conformal-prediction measures to put a calibrated number on each prediction and bring down false negatives.

Reproducing expert comparable selection

Self-directed

I framed real-estate comparable selection as learning-to-rank over roughly six million properties resolved across several databases, trained on the comp sets expert analysts had actually chosen. The result I am proudest of is a negative one. A plain per-row classifier beat the listwise ranking objectives on realistic full-pool retrieval, even though those rankers won on closed-set cross-validation. That gap turned out to be a methodological trap worth fixing: I rebuilt evaluation around full-pool tests, pre-registered ship-or-reject gates, and honest reporting of nulls, then showed the real ceiling lives in candidate generation rather than in the ranker.

Pivotal-trial biostatistics

Brought in as an independent consultant, I owned the analytics for a medical-device company's pivotal Phase III trial end to end, as the only data scientist on the engagement. I built the pipeline from locked clinical-data exports to analysis-ready datasets, then ran the statistics on top of it. Survival analysis, recurrent-event modeling, and a hierarchical composite-endpoint adjudication for win-ratio comparison. I wrote a Monte Carlo trial simulator to stress-test endpoint definitions and statistical power, and built the operational analytics that tracked activation across dozens of international trial sites. One consultant, the full statistical stack of a pivotal trial.

Maritime document intelligence

A top contributor to an Azure pipeline that turns messy maritime documents into structured, queryable records, I built several of its components. Most notably, an in-process TF-IDF and logistic-regression classifier with confidence-gated LLM escalation beat the LLM outright on the highest-stakes class, at a fraction of the cost. The last few percent of accuracy never came from more data. They came from methods that could absorb the inconsistencies of real documents.

A real-estate ML portfolio, built solo

At Starwood, every project I take on is mine end to end, built solo. Over the past year that has meant five production-grade systems: the comparable-selection research above; a data-only valuation model with rolling-origin validation and conformal intervals; a similarity engine that reconciles two databases that agree on almost nothing; a document-extraction platform now in production; and a market-scoring agent on a point-in-time panel with a deterministic, auditable rubric. What stands out is the pace: in a single recent month I shipped two of them, pushed a third to production, and built a fourth from nothing to a working v1.

conformal-selective-uq-public

A selective conformalized-quantile regression method that holds group-conditional coverage under distribution shift and abstains where the signal is genuinely unrecoverable. The abstention is validated, not asserted: the target carries a known, injected noise floor, so I can show it drops the points that were never recoverable. It makes no alpha or return claim.

quant-strategy-validation

A leakage-free validation harness for systematic-strategy research. It pre-registers its gates, walks forward with an embargo, bootstraps on event clusters, deflates the Sharpe for the trials tried, and reports its nulls honestly. It makes no alpha claims; the deliverable is the discipline to trust a negative result rather than manufacture an edge.

property-price-avm

A time-aware automated valuation model with rolling-origin validation and conformal prediction intervals. It puts calibrated uncertainty on a heavy-tailed, regime-shifting target, controls for leakage, and reports honestly on where the real ceiling lives.

comparable-property-ranker

Cross-sectional learning-to-rank for real-estate comparables, the same mechanic that drives systematic-equity selection. It is evaluated on realistic full-pool retrieval, the setup that exposes the closed-set cross-validation trap.