RESEARCH INTERESTS

Governance, explainability, and verification of agentic AI systems in regulated environments; NLP, deep learning, and privacy-preserving architectures.

---

APPLIED RESEARCH & SELECTED SYSTEMS BUILT

Cybersecurity Analysis & Documentation AI Agent 2026
Gradient Descent Labs | Developed independently and provided to the National Endowment for the Arts for agency use

• Designed and built a purpose-built agentic AI system integrating NIST 800-53 rev 5, OWASP Top 10 for 2025, MITRE ATT&CK, MITRE CVEs, and CISA KEV
• Repository scanning, vulnerability analysis, and MITRE ATT&CK assessment
• Anchors published research on purpose-built vs. large-scale model thesis (Forbes Technology Council, 2026)

Section 508 Analysis & Documentation AI Agent 2026
Gradient Descent Labs | Developed independently and provided to the National Endowment for the Arts for agency use

• Built a small, task-specific LLM agent for automated accessibility compliance evaluation
• Demonstrated applied thesis: purpose-built models outperform generalist LLMs on constrained, high-frequency tasks

GitHub README Generator — Purpose-Built Agentic System 2026
Gradient Descent Labs | Production deployment

• Deployed small, task-specific LLM agent for automated developer documentation generation
• Demonstrated applied thesis: purpose-built models outperform generalist LLMs on constrained, high-frequency tasks

DevtoDeployment — Multi-Agent Python Application 2026
Gradient Descent Labs | GCP: Cloud Run, Pub/Sub, Firestore, GCS, Secret Manager

• Architected multi-agent orchestration system using Python 3.12, LangGraph, and Google Cloud Platform
• Integrated OpenClaw agentic gateway via Telegram; Anthropic Claude API as primary LLM backend

Tsunami Prediction System — End-to-End ML Pipeline on Azure 2026
Independent Applied ML Project | github.com/jtunnessen/tsunami-prediction-system | Azure App Service

• Architected and deployed an end-to-end ML classification pipeline on Azure, optimizing recall and model interpretability for seismic risk analysis (89% tsunami recall)
• Owned the full ML lifecycle — data ingestion, automated preprocessing, hyperparameter tuning, model serialization, and cloud deployment behind a web interface for real-time prediction

Mental Health Analysis via Textual Analytics 2025
Research Project — George Washington University | GCP GPU Clusters

• Built a two-layer bidirectional LSTM with an additive attention mechanism over 300-dimensional GloVe embeddings, trained in two phases to classify self-reported mental-health discourse from a public research corpus
• Rendered token-level attention heatmaps to show which spans drove each prediction, and examined algorithmic bias, explainability, and fairness using PyTorch on Google Cloud GPU infrastructure
• Benchmarked against baseline NLP models. Academic coursework on a publicly available research dataset — not a screening instrument, not clinically validated, and not deployed

Ipsum NLP Auto-Transcription & Translation Tool 2019
Voice of America | USAGM

• Designed and deployed “Ipsum,” VOA’s internal NLP transcription and translation tool, achieving 10x speed improvement
• Adopted by four international media organizations; scaled across 20+ language services serving 275M+ weekly viewers

Whole Genome Sequencing Cloud Re-Architecture 2016
USDA Food Safety and Inspection Service | Cloud Shared Service with FDA & CDC

• Led cloud migration and re-engineering of WGS infrastructure, yielding $12M per year in cost savings
• Established first federal cloud shared service for genomic data spanning USDA, FDA, and CDC

Code for selected projects is published at github.com/JTunnessen.