Research

Agent evaluation, selective risk, and multimodal neural decoding.

Three 2026 preprints and earlier journal work. Two of the preprints ship public, reproducible code; where a number here came from a benchmark, the benchmark is linked.

The record in four numbers

3
2026 preprints — two shipping public, reproducible code
14,750
agent execution traces in AgentProp-Bench
13,859
document fields audited for selective risk
84.8%
fused EEG–fMRI decoding accuracy, within-subject

Preprints — 2026

The two arXiv papers release the harness that produced their numbers, so the figures are reproducible rather than reported. The third is under review as the benchmark companion to the first.

The complementarity claim

EEG knows when. fMRI knows where. Fused, they beat either alone.

M.S. thesis

A Multimodal Neuroimaging Method for the Prediction of Visual Stimuli

University of Cincinnati · M.S. Computer Science · GPA 3.95/4.0 · Advisor: Prof. Vikram Ravindra

EEG–fMRI fusion framework using temporal convolutional networks for neural decoding. 70-channel EEG + 3T fMRI (Wakeman–Henson dataset); SSS, ICA/wavelet-ICA, and DWT preprocessing; MNI-normalized fusiform/occipital ROIs. The combined model reached 84.8% within-subject and 81.1% cross-subject (LOSO) accuracy with ROC-AUC 0.93/0.90 — against 65.5% EEG-only and 74.6% fMRI-only, beating GRU, LSTM-CNN, and SVM baselines.

EEG · 70 CHANNELSmilliseconds — precisewhere — poorly localisedfMRI · 3Tmillimetres — localisedwhen — seconds, smearedWakeman–Henson datasetSSSICA / wavelet-ICADWTMNI-normalisedfusiform ROIoccipital ROITemporal conv. netfusionWITHIN-SUBJECT DECODING ACCURACY50%60%70%80%90%EEG only65.5%fMRI only74.6%Fused (TCN)84.8%Cross-subject (LOSO): 81.1%ROC-AUC 0.93 within / 0.90 crossAhead of GRU, LSTM-CNN and SVM baselines.Axis starts at 50% — the whole range is labelled.
The two modalities fail in opposite directions: EEG knows when to the millisecond but not where; fMRI knows where to the millimetre but smears when across seconds. Fusing them under a temporal convolutional network beat either alone by a wide margin — 84.8% against 74.6% and 65.5% — and held up leave-one-subject-out at 81.1%, which is the number that matters for whether it generalises past the people it was trained on.

Full case study — method, baselines, and what broke first →

Reproducible artifacts

Both 2026 arXiv preprints ship their harness. verifydoc is seed-pinned and regression-gated under Apache-2.0; agenthallu-bench publishes its code and data.

  • Research artifact · arXiv:2608.14639

    verifydoc

    arXiv preprint · 14 pages

    Reference implementation for “Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays.” It shows that the natural per-field accept/review procedure silently violates its own selective-risk guarantee on real documents.

    13,859
    fields evaluated
    800
    CORD receipts
    3
    failure modes diagnosed
    49.0%
    base field accuracy

    case study →↗ source↗ paper

  • Research artifact · arXiv:2604.16706

    agenthallu-bench

    arXiv preprint · 11 pages, 4 figures, 8 tables

    Code and data for “Auditing Automated Evaluation, Error Propagation, and Runtime Mitigation in Tool-Using Language Agents.” It audits the widely-assumed reliability of automated evaluation for tool-using LLM agents against human annotation.

    14,750
    execution traces
    13
    LLM agents audited
    4
    task domains
    9 + 4
    proprietary + open-weight models

    case study →↗ source↗ paper

Journal papers — 2021

Earlier co-authored work, from the undergraduate years, listed for completeness. No public identifier is quoted for these beyond the volume and page range the record holds.

Keep reading

The engineering behind these →Experience & teaching →↗ Google Scholar