Deep Learning for Multivariate Well Anomaly Detection (SPE ATCE 2026)

Designed and implemented supervised Transformer and Temporal Fusion Transformer architectures for time-series anomaly detection on multivariate well sensor data. Benchmarked attention-based models against CNN classifiers and pretrained diffusion models, demonstrating that architecture choice must align with anomaly dynamics: local convolution excels at rapid-onset events while global attention captures long-duration degradation patterns. Findings contributed to an accepted SPE ATCE 2026 paper.

Skills: Transformers, TFT, CNN, sequence classification, anomaly detection, PyTorch Ref: SPE ATCE 2026 (abstract submitted, October 2026, Houston)

Evaluation of NVIDIA NV-Tesseract-AD Foundation Model

Led hands-on technical evaluation of NVIDIA’s time-series foundation models across four model iterations (transformer-based, diffusion-based, zero-shot forecasting, and context-enhanced retrieval), serving as primary technical contact in direct collaboration with NVIDIA’s model engineers via Cognite. Identified structural limitations of forecast-error-based detection on gradual and periodic anomalies, re-implemented vendor architectures as custom multivariate alternatives trained on domain data, and reported bugs contributing to model improvement. Results adopted by Cognite’s Dune team for their application demo at NVIDIA GTC 2025.

Skills: Foundation model evaluation, zero-shot inference, diffusion models, time-series forecasting, benchmarking, Python Ref: NVIDIA NV-Tesseract-AD

Physics-Informed Feature Engineering for Equipment & Well Monitoring

Developed thermodynamic and hydraulic features that compress complex multivariate sensor data into interpretable, domain-meaningful signals for anomaly detection. Designed U-ratio (thermal efficiency tracking) for heat exchanger fouling detection, dP/Q (hydraulic resistance) for well tubing health monitoring, and production index frameworks for reservoir productivity diagnostics. These physics-informed representations significantly outperform raw signal inputs in both model robustness and interpretability, and form the basis for production monitoring use cases scoped under the Yggdrasil platform.

Skills: Physics-informed modeling, thermodynamics, fluid mechanics, feature engineering, time-series analysis, Python

Data-Driven Tortuosity: Drilling Path Quality Prediction (PoC)

Designed a two-layer ML approach to predict micro-tortuosity risk from high-frequency drilling signals and quantify its impact on non-productive time (NPT). Developed the analytical framework linking drilling dynamics (torque, drag, vibration patterns) to fine-scale wellbore geometry quality, addressing a gap where standard directional surveys miss critical inter-station deviations. Includes early-warning indicators for stuck pipe, wiper trips, and casing difficulties derived from progressive symptom pattern mining.

Skills: Time-series ML, drilling engineering, feature engineering, predictive modeling, Python

CNN-Based Well Anomaly Detection

Developed a physics-informed deep learning pipeline for anomaly detection on high-frequency well and sensor time-series data. Designed CNN-based classifiers and signal-processing workflows to detect degradation patterns, transient flow abnormalities, and equipment malfunction signatures.

Skills: CNNs, time-series ML, physics-informed modeling, Python, Azure ML

Machine Learning for Mortgage Risk: An Interpretable Approach (PNC Bank)

Developed interpretable ML models to assess climate-driven mortgage default risk under macroeconomic and weather uncertainty. Applied SHAP-based explainability to support transparent and decision-relevant risk analysis.

Skills: LSTM, Temporal Fusion Transformer, time-series modeling, interpretable ML, risk analytics, Python

Application of Convolutional Neural Network to Predict Airfoil Lift Coefficient

Early deep learning surrogate model integrating physics-informed constraints to predict aerodynamic coefficients with high accuracy.

Skills: CNNs, surrogate modeling, aerodynamics Ref: AIAA 2018

Learning Aerodynamics with Neural Networks

Developed a Custom Element Spatial CNN (ESCNN) using coordinate-based inputs to predict aerodynamic behavior with reduced model complexity.

Skills: Data representation, physics-aware ML, deep learning Ref: AIAA 2021, Scientific Reports 2022

DockingNet: Deep Neural Network for Airfoil Optimization

Hybrid generative–predictive neural network for accelerating airfoil design optimization without CFD-based iteration.

Skills: Generative modeling, optimization, lift/drag estimation Ref: AIAA 2022

Dedalus: AI-Powered Inventory Assistant (In Progress)

NLP-enabled assistant for querying fragmented inventory and test data across Excel, Access, and PDF sources; migrated backend to SQLite for structured access and downstream analytics.

Skills: NLP, LLMs, data integration, Python, SQLite

Electrochemical Model for Dissolvable Alloys

Hybrid physics–data model predicting dissolution rates under realistic operating conditions using calibrated regression.

Skills: Multiphysics modeling, regression, electrochemistry

Hemodynamics Surrogate Model (Medical AI)

UNet-based deep learning surrogate model predicting stenotic blood flow fields in seconds from CFD-trained data.

Skills: Deep learning, UNet, biomedical CFD, Python

Petit Appétit: Weekly Meal & Grocery Planner

Web-based pipeline for meal planning and grocery list generation using Dask, AWS S3, and web scraping; planned NLP-based extensions.

Skills: Data pipelines, scraping, Dask, Python, Django

Deep Learning Chess AI (Exploratory)

Self-learning chess AI trained using reinforcement learning and neural policy optimization through self-play.

Skills: Reinforcement learning, TensorFlow, policy optimization

COVID-19 and Education Completion Analysis

Global data analysis using UNESCO and World Bank datasets to quantify the impact of COVID-19 on education outcomes.

Skills: Data imputation, regression analysis, clustering, policy-oriented analytics