My doctoral research concerns when a language classifier is adequate for a specific decision. I also build research software for agentic systems, verification, and scientific data.
Doctoral research in progress
Advisor: Dr. Jeanna Matthews, Professor of Computer Science, Clarkson University.
Member, Explainability & Transparency Group, Clarkson University, since September 2026.
My doctoral research, advised by Dr. Jeanna Matthews at Clarkson University, asks when a text classifier is adequate for a specific decision rather than simply best on a benchmark. I am developing a consequence-driven framework that starts from what an error costs the people who act on a model’s output. It converts those consequences into operating requirements such as recall floors, false-positive limits and calibration, then qualifies and selects candidate models against them with explicit uncertainty bounds. The worked case is detecting subjectively biased wording in Wikipedia sentences to support human editors, using the WIKIBIAS benchmark. The framework and study protocol are in progress; no evaluation results are claimed yet.
The study compares fine-tuned encoder classifiers with small generative language models, including whether added inference-time reasoning is worth its cost, and examines calibration and deferral.
Model evaluation and selection, uncertainty quantification and calibration, and NLP classification.
Human-in-the-loop decision support, selective prediction and deferral to human reviewers, and responsible AI.
Model orchestration, structured tool use, persistent state, agent memory, and verification workflows.
Earlier research applications include biometric measurement, computer vision, sensor fusion, and GNSS-denied autonomous navigation.
Source code, documentation, and evaluation workflows.
Shared memory and review workflows for research teams and AI assistants.
GitHub-native collaboration with a built-in knowledge graph and 22 MCP tools. Tracks artifacts, review gates, context packs, and distributed tasks using git as the shared substrate.
Face-image-quality measurement with a reference comparison workflow.
Python implementation of the fixed BSI OFIQ v1.1.0 profile: 27 quality components plus the unified quality score. Includes configuration and model integrity checks and a live C++ comparison workflow.
Validation that checks what a knowledge bundle means, as well as its format.
Tools to validate, lint, migrate, and index Open Knowledge Format bundles. Detects broken relationships and silent semantic failures that basic conformance checks do not reject.
Lunar terrain authoring grounded in real south-polar terrain products.
Terrain generation, authoring, and preprocessing for Godot robotics simulation. Uses LOLA/PGDA terrain products, ephemeris-derived illumination, explicit provenance, automated validation, and CI.
A common evaluation interface for SwinFace and ArcFace encoders.
Reproducible face verification over public benchmark data, with documented provisioning, integrity checks, metrics, limitations, and optional mixed-effects analysis.
Knowledge-graph and geospatial exploration of Medicaid provider spending.
An exploration system for a 227-million-row HHS T-MSIS dataset. Combines DuckDB, NetworkX, FastAPI, React, MapLibre/deck.gl, and agent-assisted querying.
Design stage
A scientific IDE for biometric quality, verification research, controlled degradation experiments, and publication-artifact generation. The public work currently comprises product architecture, scientific schemas, and decision records.
Read the architecture and design records