Efficient and Reliable Machine Learning
Data Science, University of Wisconsin-Madison
I study how to make machine-learning systems more efficient without losing the computations and representations that matter for their behavior. My work spans numerical precision, inverse vision, neural operators, and mechanistic interpretability. I also contribute to diffusion-based reconstruction research in Professor Grigoris Chrysos's group at UW-Madison.
Research Résumé · Academic CV · LinkedIn · Hugging Face · Email
Which computations can we approximate or remove without changing the capabilities we need to preserve?
I investigate precision-sensitive operations, model internals, and evaluation methods that reveal when efficiency gains come at a meaningful cost to reliability.
The Measurement Operator Changes Which Bits Matter: Operator-Conditioned Mixed Precision for Inverse Vision
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027), Algorithms Track. Accepted.
Introduces Operator-Susceptibility Matching (OSM), which uses information about the sensing operator to select mixed-precision schedules. In a controlled eight-stage study, OSM reduced mean normalized operator-visible regret to 2.7%, compared with 6.3% for the best global-static schedule at the same three-stage high-precision budget. A separate native A100 timing experiment recorded 16.4 ms per reconstruction versus 23.6 ms for its FP16 reference (1.44× speedup). The quality and timing results use distinct evaluation protocols.
EvidenceLens: A Declarative Evidence-State Interface for Human-AI Visual Inspection
ACM UIST Adjunct 2026, poster publication.
A declarative evidence-state interface for bounded human-AI visual inspection and review.
Publication DOI · Paper · Repository
Quantized Defect Correction for Fractional Neural Operators
Investigates precision-sensitive failures in fractional-PDE neural operators through activation and weight quantization, physics-residual diagnostics, and selective defect correction with numerical and hardware evaluations.
Preprint · Code and reproduction · Software DOI
EvAudit: Latent Proxy-Evidence Auditing for Generative Data Mining
Developed an evidence-auditing framework to filter unsupported claims before downstream use. The public research artifact includes calibration and stability checks, validation utilities, reproducibility scripts, and a 2026 preprint.
Preprint and research artifact · Software DOI
Multilingual LLM Mechanistic Interpretability
Contributed to Qwen and Llama circuit-discovery studies using activation patching, attention-path interventions, and held-out controls. Additional experiments examine whether language-control mechanisms transfer across model variants.
Qwen circuit study · Llama validation · Cross-model analysis
-
Station House, LLM Consciousness Self-Attribution: Developed OLMo 3 identity-spoofing elicitation pilots across SFT, DPO, and Instruct, covering 81 recorded conversations across three training stages. Contributed evaluation tooling and grading-context analysis. Initial pilot · Training-stage extension · Evaluation tooling
-
Station House, Marin Neuron Hunt: Investigated Marin 32B neuron behavior using activation examples and targeted contrast experiments. Developed and revised neuron explanations through negative-side tests, numeric-context sensitivity checks, and follow-up experiments that refuted earlier hypotheses. Multiple contributions were merged. Initial explanations · Verification experiments · Numeric-context tests · Revised explanation
-
Station House, ConspiracyBench: Contributed a merged, evidence-backed benchmark question and examined benchmark design, duplicate detection, evidence-focused evaluation, and paraphrase robustness. Merged benchmark contribution · Evaluation proposal
-
Medical-Imaging Robustness Study: Contributed merged model-training, inference, and evaluation infrastructure, including dataset validation, baseline setup, prediction exports, and an ImageNet-pretrained baseline. Dataset and training setup · Prediction exports · Baseline implementation
-
StellarResearch, AdaptiveRL: Added a merged LiDAR noise and dropout robustness benchmark for an adaptive reinforcement-learning project. Robustness benchmark
-
AladdinAI: Contributed merged session-aware dataset splitting and DOCX parsing fixes. The inline-image ordering fix was recognized by the XERJ Corpus Hub. Dataset splitting · DOCX fix · Contributor page
-
CausationEntropy: Contributed conditional mutual information estimator fixes and Gaussian CMI performance optimization. CMI fix · Optimization
- UW-Madison Sustainability Symposium 2026: When Does Better AI Actually Save More Energy? Decision-Aware Adaptive Fidelity for Building Control. Poster · DOI
- WiSys SPARK Symposium 2026: Placement-Aware Quantization for Fractional-PDE Neural Surrogates. DOI
- Pi Mu Epsilon Regional Undergraduate Mathematics Conference 2025: Numerical Stability and Convergence of Physics-Constrained Neural Networks in Vibration Modeling. DOI
- Judge's Choice Award, Most Innovative Track (2026): UW-Madison Tech Exploration Lab Demo Night, PawSight.
- 1st Place, Culver's Business Model Competition (2025): UW Oshkosh, CurbMap 360. University feature.
- Calculus Prize Exam Winner (2025): $1,000 mathematics scholarship.
More experimental work, preprints, and software are available in my research repositories.