Ricky Gole 路 Baltimore, MD

I find the failure modes ML systems hide after deployment.

Graduate researcher in Advanced Computing at Morgan State (4.0 GPA), working on evaluation and robustness of LLM and clinical ML systems. My work is about the gap between benchmark accuracy and behavior under real distribution shift, and about building the audits that close it. PyTorch, LLM fine-tuning, benchmark design.

Graduating December 2026 路 available full-time January 2027 Authorized to work in the US Remote or relocation

Selected findings

Each one is a failure nobody was measuring, with a fix attached and, where the code is public, a repo you can run.

Temporal OOD audit 路 clinical NLP

A model that scores 0.73, then zero

0.73 → 0.00
BiomedBERT F1 on post-2019 FDA drug interactions

A drug interaction classifier holding 0.7273 F1 in distribution collapsed to exactly zero on a 600-pair post-2019 out-of-distribution set, reproduced across four seeds. It did not degrade. It flatlined. A name-only in-distribution control reproduced the same 0.0000, which pins the cause on structural dependence on annotated sentence context, not drug novelty. I recovered 0.2640 F1 with KnowledgeRAG-DDI, a hybrid BM25 and FAISS retrieval probe over FDA labels with leakage-controlled indexing and reciprocal rank fusion.

IEEE ICDM 2026 路 accepted

Five models, wrong in unison

19.9×
more unanimous-wrong answers than chance predicts

Across 4,089 USMLE and MedMCQA questions, five models (Llama-3-8B, Qwen-2.5-7B, Gemma-3n, GPT-4o, DeepSeek-V3) agreed on a wrong answer 24.6% of the time against a 1.23% chance baseline. A distractor shuffle isolated the cause as content driven, not positional. The resulting parameter-free routing rule flags these cases at 83.0 to 96.1% precision (96.1% on MedQA, 83.0% on MedMCQA) for 6.7 to 8.0% human review cost, holding 77.5% under a held-out frontier-model substitution.

Ensembles don't average out error. They concentrate it. llm-convergent-failures
Fairness benchmark 路 under review, EMNLP

DART: dialect robustness in essay scoring

480 pairs
six U.S. English dialect families, plus 160 controls

A controlled benchmark validated by a three-gate pipeline (length drift, DeBERTa-v3 NLI entailment, regex feature presence) at a 92.71% pass rate. Mixed-effects regression exposed directional bias against African American English in a fine-tuned scorer (Pearson r = 0.829, QWK = 0.714) as the only family with significant negative drift (β = −0.029, p < 0.001), concentrated at the equalised-odds decision threshold. Chain-of-thought scaffolding amplified the drift across three LLM judges rather than reducing it.

Multi-omic fusion 路 real patient cohort

SGOF: when more data makes it worse

−0.149 → 0.012
AUROC drop under an adversarial modality: MOGONET vs. SGOF

On a real 111-patient antidepressant response cohort, adding an uninformative 258,782-SNP genotype layer dragged four fusion families below the best single modality (methylation alone, AUROC 0.692), collapsing naive concatenation to 0.334. I attributed it to uninformativeness with a synthetic noise control, then built Signal-Gated Omic Fusion, which weights modalities by held-out validation signal and held at 0.664 (paired-bootstrap change 0.012, 95% CI [−0.035, 0.059], not significant) where MOGONET dropped 0.149. The claim is stability under a bad modality, not beating the best single one.

Shipped

Working software, not only papers. Everything here is public and runnable.

overturn

An agent fleet that drafts and tracks health insurance denial appeals across a multi-week lifecycle.

Python

Bias-Lab

An interactive lab where students train a classifier in the browser, drag one threshold, and watch who absorbs the error.

JavaScript

k2l-readaloud

Browser-only oral reading fluency assessor. Local Whisper plus a Needleman-Wunsch running record.

JavaScript

semantic-gap-educational-rag

Code for the EDM 2026 paper on behavioral embeddings and why linear methods fail for educational RAG.

Python

sesame-robot

Custom ESP32-WROOM firmware for a quadruped robot. IK gaits, closed-loop sensing, LLM-driven behavior.

C++

AItailoring

In-browser AI resume tailor and ATS optimizer powered by WebLLM. Runs entirely client side.

JavaScript

Publications

Accepted and published 5

IEEE ICDM 2026

Detecting and Mitigating Convergent Failures in Clinical LLM Ensembles Gole, R., et al. First author 路 code (llm-convergent-failures)

ACM L@S 2026

Prescriptive Persistence: Quantifying the Breakdown in Human-AI Pedagogical Co-Regulation in ELL Writing Feedback Gole, R., et al. First author

EDM 2026

The Semantic Gap in Behavioral Embeddings: Why Linear Methods Fail for Educational RAG in Mathematics Gole, R., Dacon, J. First author 路 code (semantic-gap-educational-rag)

Cloud Summit 2026

Mitigating Confused Deputy Attacks in AWS Cross-Account IAM Roles Through ExternalId Enforcement Gole, R., Heydari, V. First author

IEEE BIBM 2025

BIO-DQNA: Meta-Learning and Contrastive Reinforcement Learning for Personalized Comorbidity Management in Type 1 Diabetes and Hypertension Dacon, J., Gole, R., et al. Second author

In submission 2

Manuscript

DART: A Controlled Benchmark for Dialect Robustness in Automated Essay Scoring

Manuscript

Beyond the Benchmark: Diagnosing Knowledge Tracing Evaluation Under Distribution Shift

Additional peer-reviewed work 4

JISE 2026

Understanding AI Literacy Among College Students: Voices from a Historically Black College and University Journal of Interdisciplinary Studies in Education

AJSE 2026

Cybersecurity and International Collaboration in STEM: The Role of Global Partnerships in Protecting Digital Infrastructures and Data American Journal of STEM Education

2026

Bridging the Digital Divide: Evaluating the Impact of Digital Literacy Training for School Students in Baltimore's Underserved Communities

2026

Trapped-Ion Qubits: Architecture, Implementation, and the Road to Fault-Tolerant Quantum Computing

Full record and citation counts on Google Scholar.

Also on the bench

Benchmark validity

35.7% of a deep model's edge, from two features

Across 95.3M EdNet interactions, a two-feature baseline recovered 35.7% of Deep Knowledge Tracing's above-chance AUC, a construct-validity confound. I introduced the recovery ratio as a reusable diagnostic, with calibration reporting (Brier, ECE, log-loss) and multi-seed stability checks.

If a trivial baseline gets a third of your gain, the metric is the problem.
Safety-constrained RL 路 BIO-DQNA, second author

Unsafe actions, 12% to 2%

Clinically tiered reward functions on a custom PyTorch environment raised policy success to 94% while cutting the unsafe-action rate from 12% to 2%. Benchmarked six DQN variants (Standard, Double, Dueling, Categorical, Noisy, Rainbow) at 100 seeds each with GPU hyperparameter sweeps. Published at IEEE BIBM 2025 with J. Dacon as first author.

LLM evaluation

Over-correction, 42.9% to 6.6%

Quantified AI and human pedagogical divergence across 12,979 ELL sentences and introduced the Prescriptive Gap metric (t = 103.19, p < 0.001, d = 0.91). A controlled frugal-prompt intervention cut GPT-5.2 over-correction from 42.9% to 6.6%.

Cloud security

Closing the confused deputy

Blocked cross-account confused-deputy attacks across six scenarios using paired IAM trust policies (ExternalId, SourceAccount, role enumeration), and processed 1M audit-log events through an entropy-based anomaly detection pipeline over STS AssumeRole requests.

Toolkit

Machine learning

LLM fine-tuning (DPO, LoRA)RAG TransformersReinforcement learning Rainbow / Dueling DQNContrastive learning Meta-learningNLP

Evaluation and methods

Benchmark designDistribution-shift audits Ablation studiesBootstrap CIs Mixed-effects regressionCalibration (Brier, ECE) Hyperparameter sweepsMulti-seed stability

Engineering

PythonPyTorchSQL JavaC++ HuggingFacescikit-learn Pandas / NumPy

Systems and data

GPU training infraTraining pipelines Hybrid retrieval (BM25, FAISS)ETL AWS (IAM, STS, CloudTrail)Linux GitJupyter

Production systems and reliability

Four years keeping other people's systems up, before and alongside the research.

2022

Level 2 Technology Consultant, TeamLogicIT

Monitored 50+ enterprise networks in a SOC, cutting response time 30%. Deployed Sophos firewalls, Cisco Duo MFA, and Active Directory controls, reducing unauthorized access 35%. Documented SOX and SOC 2 findings.

2019 to 2023

Senior Student Technician, Caldwell University

Sustained 99% uptime across VMware VDI, Linux, and Windows for 500+ users. Cut recurring tickets 20% by automating log analysis and health checks in Python.

2025 to present

Teaching Assistant, Morgan State University

Supported 50+ students across operating systems and hardware courses. Lab instruction and debugging in Linux, Windows, and systems programming.

Education

2025 to Dec 2026

M.S. Advanced Computing, Morgan State University

GPA 4.0/4.0. Graduate researcher. Deep cloud learning, cybersecurity, data science, network security, artificial intelligence.

2019 to 2023

B.S. Computer Information Systems and Financial Economics, Caldwell University

GPA 3.55/4.0. Software engineering, statistics, operational management, data disaster recovery.

Get in touch

If your team cares whether its models hold up in production, and not only whether they win on a benchmark, I would like to talk.