Projects & Publications

Clinical foundation model initiative on the Johns Hopkins REACH de-identified EHR dataset for congestive heart failure readmission and mortality prediction.

• Profiled 8.3M patients (3.2M with clinical notes, 1.12B structured OMOP events) and designed the CHF cohort: 103,577 patients, 17.1% 30-day readmission, 28.8% mortality
• Developed a staged 4-phase IRB access strategy (single-condition retrospective through generalist CDS), escalated to school leadership
• Starting with fine-tuned clinical BERT encoder for binary and time-to-event readmission classification, with an autoregressive rollout planned for reasoning traces

AI-powered knowledge platform enabling faculty and staff to query HR policy documents with cited, sourced answers via a RAG pipeline.

• Built end-to-end: PDF ingestion with Claude Haiku Markdown reformatting, chunked full-text indexing in PostgreSQL, source-cited chat interface, interactive document viewer with passage highlighting
• Deployed on Azure App Service with React/TypeScript frontend and Bun/Elysia backend, using Azure Blob Storage and LLM tool-calling (GPT + Claude)
• Shipped production enhancements covering highlight precision, Azure SAS credential fallbacks, deduplication, and PDF text normalization for legacy documents

Automated system evaluating faculty CVs against AACSB accreditation standards (SA/PA/SP/IP qualification types) for the Carey Business School.

• Multi-stage LLM pipeline: concurrent chunk extraction, activity enrichment, and verdict evaluation across a 6-year rolling academic window
• Built a consistency harness to validate determinism across repeated runs
• Refactored domain rules into standalone config files, making yearly academic calendar updates a config change instead of a code change

Multi-modal pipeline for automated paper quality screening, combining visual analysis and full-text evaluation into a normalized quality score.

• Fuses first-page visual analysis, extracted figure inspection, and full-text evaluation across 35 weighted rubric criteria (visual, stylistic, content)
• Evaluated on 200 papers: 0 false negatives on high-quality papers, 5/7 low-quality papers correctly flagged
• Benchmarking open-source alternatives with Qwen2.5-VL-7B served via vLLM under identical rubric and scoring scheme

Evaluated tampering detection capabilities of 120B-235B parameter reasoning models by systematically perturbing chain-of-thought traces across multi-domain reasoning benchmarks.

• Built an Inspect AI evaluation pipeline applying step deletion, cross-model substitution, and semantic injection
• Quantified detection performance using nonparametric bootstrap confidence intervals
• Found that detection sensitivity differs fundamentally between reasoning traces and final outputs

Agentic RAG system for structured financial reasoning over 83 Canadian earnings call transcripts across 13 major equities spanning Q2 2024 - Q1 2026.

• Built a LangGraph agent with routing, retrieval grading, hallucination detection, and self-correcting reformulation loops
• Integrated PostgreSQL + pgvector with 768-D local embeddings; replaced LDA with BERTopic for stable topic discovery
• Speaker-aware chunking enabling management-vs-analyst sentiment gap analysis with FinBERT

Corpus-reasoning system answering natural-language questions across 99 ML papers on maternal morbidity, combining symbolic reasoning, citation-graph traversal, SQL aggregation, and semantic retrieval.

• Hybrid architecture: SQLite, FAISS, and NetworkX citation graphs indexing ~12,000 claim units
• Typed query routing across eight reasoning tiers including contradiction detection and quantitative reasoning
• Automated extraction from 99 papers using Anthropic Haiku at sub-$0.002 average query cost

Credit default prediction using regularized logistic regression with temporal stability, asymmetric error costs, and decision-aware calibration.

• Behaviorally motivated features from utilization, payment history, and exposure dynamics
• AUC 0.82 and KS 0.41, improving rank-ordering power by 17% over baseline
• Reduced false negatives by 24% at fixed precision under cost-sensitive objectives

Production ransomware detection using ensemble classifiers trained on 500,000+ FlashCore Module filesystem traces, integrated into the Storage Virtualize stack.

• Trained SnapML-based ensemble models with automated class re-balancing and cross-family validation
• Achieved <1% false-positive rate on live production workloads across unseen ransomware variants
• Integrated inference to trigger early alerts, immutable snapshots, and forensic trace dashboards

Forecasting-driven placement system for enterprise block-storage I/O demand prediction and SLA-aware optimization across 100+ FlashSystem arrays.

• Benchmarked ARIMA against PatchTST and Tiny Time Mixer (TTM); improved out-of-sample accuracy by 25%
• Operationalized a decision layer translating probabilistic forecasts into placement and migration actions
• Prevented nearly 10,000 potential SLA violations through forecast-aware scoring

Open-source Model Context Protocol server exposing IBM Storage Insights observability capabilities to agent-based AI workflows and Claude Desktop.

• Python MCP server exposing 13 observability tools backed by Storage Insights APIs
• stdio and streamable HTTP transport with multi-tenant credential handling
• Reduced end-to-end issue identification latency by 15% through autonomous agent-driven diagnostics

GDPR-safe support-ticket deflection pipeline using clustering and retrieval-driven FAQ generation across 26,000+ historical tickets.

• Sanitized GDPR-sensitive fields and clustered tickets into representative knowledge groups
• Auto-generated validated FAQs with grounding checks to prevent hallucinated answers
• Reduced Level 2 support workload by ~30% and mean time to resolution by 20%

Natural-language observability chatbot translating free-text queries into structured Storage Insights API calls using Granite and Llama models via IBM Watsonx.

• Intent detection and entity extraction for alerts, metrics, volumes, and tenant diagnostics
• Community-driven reference architecture for GenAI-enabled AIOps workflows

Sparse Mixture-of-Experts Transformer for multilingual language modeling with top-k routing and load-balancing loss in PyTorch.

• Scaled parameter count by 3x at constant inference FLOPs via sparse expert routing
• Analyzed token-level expert specialization across multiple languages

How linguistic framing affects reasoning reliability and token usage in Llama 3, GPT-4, and Claude Sonnet.

• Evaluated controlled prompt variant sets across multiple frontier models
• Minimal accuracy variance but measurably higher token usage under emphatic framing
• Validated with McNemar tests, paired t-tests, and effect-size analysis

Ensemble-based predictive modeling and genetic algorithm optimization for team construction under budget and role constraints.

• Ensemble models estimating performance and expected value under uncertainty
• GA framework optimizing budget, role diversity, and risk-adjusted returns
• 61% accuracy identifying top-performing players; published in Springer LNNS