Research questions turned into working systems.

Reliable Tool-Using Agents

Question. How can runtime defenses block unsafe tool actions while preserving useful task completion—and how should recovery be evaluated after an intervention?

Contribution. Designed evaluation of safety, task completion, structured block feedback, and post-intervention recovery; integrated online action interception and performed leakage auditing; co-developed the benchmark, evaluation pipeline, and closed-loop trace-replay studies.

Ongoing collaborative research. No publication status, private research artifact, or sole-ownership claim is implied.

New Haven Spatial Analysis

Question. What spatial patterns appear across 331,423 estimated building-height records, and how can a focused 16,379-polygon preview make the work explorable in a browser?

Contribution. Exploratory spatial statistics, Global and Local Moran’s I, a spatial-lag experiment, interactive polygons, and a 3D view.

Open case study →

Individual course case study. UT-GLOBUS values are estimates, not surveyed ground truth.

Evidence-Governed Career Operations

Problem. Job workflows combine changing listings, personal evidence, employer-level conflicts, and actions that should never happen without approval.

Contribution. Extended Career-Ops with an explicit state controller, evidence-linked résumé routing, time-bounded checks, submit leases, opportunity protection, and ATS form-state validation.

An extension of the open-source Career-Ops project by santifer. Personal applications, credentials, browser state, and operational data remain private.

NYC Rodent Inspection Analysis

Question. How do inspection outcomes vary across New York City, and are reported sightings associated with initial inspection pass rates?

Contribution. Linked public datasets, aggregated ZIP-level measures, tested associations, fit a binomial model, and built geospatial visualizations presented to the NYC Open Data team.

Personal course contribution only. No address-level records or original course/Classroom repositories are published here.

NBA Salary Prediction

Question. How much of a player’s salary can be explained by measurable on-court performance?

Contribution. Linked performance and salary data, engineered per-game features, compared four models, tuned hyperparameters, and evaluated a weighted ensemble.

Individual course analysis, not a production salary model. The original course repository remains untouched.

Machine Learning & Data Systems

Additional confirmed work includes an end-to-end RAG course pipeline, PyTorch image classification and transfer learning, Hugging Face question answering, and Spark DataFrame/SQL processing.

These are supporting examples of hands-on technical work rather than claims of production deployment.

Course and project context is stated explicitly; no unverified metrics or company affiliation are attached.