Applied Scientist & Machine Learning Engineer — LLMs, VLMs, and multimodal computer vision.

Director of the eXplainable Deep Intelligence (XDI) Lab at North Carolina A&T State University. I build and deploy production ML systems — aviation safety (Boeing), autonomous driving (GM), and clinical AI (N3C) — spanning fine-tuning, distributed training, and serving. 32 peer-reviewed publications (192 citations, h-index 7), $5M+ in funded research (NNSA, Boeing, DOE, NIH), directing 8+ researchers on a 20+ GPU cluster.

Recipient of the IEEE 2025 Outstanding AI Achievement Award for Aviation and Healthcare. Chair, IEEE Eastern North Carolina Section Computer Society.

📄 Google Scholar · 💼 LinkedIn · 🌐 xdilab.com · ✉️ hrmoradi@gmail.com

Research code lives in the lab account, not on this profile:

The tables below map each research area to its repositories.

Fine-tuned open models (LLaMA, Mistral, Qwen, Phi, Gemma) with LoRA/PEFT for pilot-query interaction and anomaly detection over multi-frequency air traffic control data. A 3-stage adaptation pipeline (domain-adaptive pretraining → SFT → instruction tuning) reduced Slot Error Rate 32% → 5%; a self-supervised post-ASR correction loop (BART, T5) learning from multi-ASR transcription errors cut Word Error Rate 16% → 8%.

LLaMA Mistral Qwen LoRA/PEFT BART T5 Hugging Face DeepSpeed

Computer-vision perception deployed on a full-scale autonomous vehicle (ROS 1 Noetic), integrated over CAN bus: YOLOv10 detection (mAP@0.5 = 0.52, BDD100K), TwinLiteNet drivable-area segmentation (mIoU = 0.91), EfficientNet-B0 traffic-sign/light classification (99.4% / 96.5%, MTSD/LISA), and a CRAFT + CRNN OCR speed-limit reader.

YOLOv10 TwinLiteNet EfficientNet CRAFT+CRNN ROS CAN bus PyTorch

Open-vocabulary detection and segmentation of privacy-sensitive objects (GroundingDINO + SAM 3.1) with trainable class-anchored soft-prompt tuning — +5% accuracy overall, up to +50% on rare classes. Extending to few-shot learning.

GroundingDINO SAM Prompt Tuning Few-Shot Learning VLMs

Curriculum learning and progressive-loss training for hip/knee radiograph segmentation — reduced demographic Skewed Error Ratio 5–10% while improving IoU 4%+ (761-patient cohort). Extended to auditing bias in foundation and vision-language models.

Fairness Metrics Curriculum Learning Segmentation PyTorch

Clinical ML on a 20M+ patient repository, cohorting to 83K admitted patients (28-day discharge/death outcomes) via PySpark in secure Palantir Foundry environments. An explainable treatment-effectiveness pipeline combining propensity-score matching with SHAP quantified drug-combination effects at 0.90 AUC / 81% accuracy (GBDT).

PySpark Palantir Foundry SHAP Propensity-Score Matching GBDT

Fine-tuned LLaMA (LoRA), BioClinicalBERT, BiomedNLP, RoBERTa, and DistilBERT for retrieval and reasoning over unstructured medical notes.

LLaMA BioClinicalBERT RoBERTa DistilBERT LoRA RAG

LLMs Topic Modeling Performance Modeling Distributed Systems C++

End-to-end personal AI assistant: a LangGraph agent with native LLM tool-calling (document retrieval over Weaviate, vision, gated email workflow), served via FastAPI, exposed through a Cloudflare Tunnel reverse proxy with API-key auth, and instrumented with a self-hosted MLflow tracking server logging per-stage latency alongside a qualitative feedback loop. Includes a human-in-the-loop safety gate — the model can only request the one irreversible action (sending email); the actual send is architecturally unreachable from its own tool-calling loop.

Python LangGraph FastAPI Weaviate Ollama MLflow Docker Whisper Kokoro TTS

• Generative AI for Air Traffic Communication Automation: Fine-Tuned Language Models Using Simulation-Grounded and Real-Trace Data — IEEE Access (under review), 2025
• Enhancing Segmentation Fairness Through Curriculum Learning and Progressive Loss: Hip and Knee Radiographs — AAAI Workshop on Health Intelligence (W3PHIAI), 2025
• Using Topic Modeling and LLMs to Recommend CAPEC Attack Patterns — IEEE Conference on Cyber Security and Resilience (CSR), 2025
• Online Performance Modeling and Prediction for Single-VM Applications in Multi-Tenant Clouds — IEEE Transactions on Cloud Computing, 2021

Full list of 32 publications on Google Scholar.

ML/DL PyTorch · Hugging Face Transformers · DeepSpeed · LoRA/PEFT/BF16 LLM/RAG LangGraph · LangChain · Weaviate · FAISS · Ollama · Instruction Tuning Vision YOLO · GroundingDINO · SAM · EfficientNet · CRAFT+CRNN · Prompt Tuning Serving & Ops FastAPI · Docker · Gradio · MLflow · Git Data PySpark · Spark/Hadoop · SQL · Palantir Foundry · AWS · GCP Explainability SHAP · XAI · Propensity-Score Matching · Fairness Metrics