AI Systems & Machine Learning

Simulation-Driven Multi-Agent Blast Planning

Production Dyno Nobel, 2024 - 2025 Led a four-person team building and productionizing a cloud-native simulation and optimization platform around patent-pending C++ physics models, deployed on AWS Fargate behind API Gateway and an ALB with full CI/CD.

On top of that platform I designed a tool-using multi-agent system that invokes the simulation services, evaluates candidate blast plans against the physics output, and generates recommendations under domain constraints, connecting model results directly to engineers' decision workflows. I also created the company's first RAG platform, from document ingestion and indexing through prompt tuning, model evaluation, API serving, and rollout, and defined its success criteria with domain experts.

Pallas14 - LLM-Driven Engineering IDE

In Development Browser demo live A local agentic engineering workstation that uses your own LLM subscription to drive the PCB design, packaging, signal-processing, and thermal-analysis tools already on your machine, taking a project from idea to manufacturing files across multi-step engineering workflows.

The core problem is the same one that shows up in every applied-AI engagement: getting a model to reliably operate real engineering software, check its own work against those tools' output, and stop when a result isn't good enough. Pallas14 is my testbed for agent harnesses, tool integrations, and verification loops around existing engineering codes rather than a new walled garden.

JEV Simulator - AI-Integrated Physics Engine

Open Source An open-source rigid-body physics-engine prototype that uses TypeSafe AI's JEV model as a learned narrow-phase collision classifier for point-cloud rigid bodies. The conventional parts of the pipeline, broad-phase detection, contact generation, impulse response, integration, and adaptive timestepping, run on CPU-only PyTorch.

The engine ships with unit and end-to-end tests and a geometric proximity detector as an offline stand-in, so the learned classifier can be swapped out and checked against a known-correct baseline.

GlimzAI - Open-Source Agentic News Platform

Live Project Open Source 2026 Preprint A fully automated LLM-agent pipeline that ingests daily news, judges which stories can become predictive questions, generates and revises polls, deduplicates them semantically with sentence embeddings, summarizes the underlying coverage, generates infographics, and distributes the result as native polls on X. No human in the loop from ingestion to publication.

Runs as scheduled AWS Lambda tasks inside a VPC with RDS PostgreSQL, EventBridge cron triggers, S3 for generated media, and Terraform-managed infrastructure deployed through GitHub Actions. The React front end is served from S3 and CloudFront. The end-to-end architecture is documented in a March 2026 preprint.

Signal Anomaly Detection: An LSTM Approach

In this project, I designed an innovative architecture that leverages a fusion of LSTM and CNN layers. The input dataset consists of multiple channels, encompassing signal data, signal variation, and a channel dedicated to frequency characteristics.

While the model was initially trained on normal signals without any anomalies, its predictive ability diverges significantly when confronted with signals exhibiting anomalous behavior. To identify these anomalous segments within the signal, I employed the Kullback-Leibler divergence, comparing the distribution of the prediction-error with that of a standard normal distribution.

A Semantic Segmentation Solution for Oil and Gas Pipeline Component Detection

I have developed a Semantic Segmentation model that represents the weighted average of four different models. I used this model to detect the components of a pipeline based on Magnetic Flux Leakage (MFL) data. To the best of my knowledge, this marks the first instance where Semantic Segmentation has been applied for this specific purpose.

This ensemble model comprises four distinct components, each built using a unique combination of architectures: FPN and Unet, alongside encoders resnext and densenet. Impressively, the final model exhibited outstanding performance, boasting a remarkable 90% recall rate and a solid 77% precision rate in detecting components. This work has been accepted to ICMLA 2023.

Pipeline Defect Characterization Using a Boosting Model

I have designed a LightGBM model to forecast the dimensions of defects, specifically their geometric properties, based on non-destructive measurement signals. The extracted features encompass a combination of signal characteristics at the defect's location and the measurement conditions.

To enhance the model's performance, I employed Bayesian Optimization via Optuna and utilized the Tree-structured Parzen Estimator (TPE) algorithm for hyperparameter tuning. The outcome of the final model showed a significant improvement in the R2-score, MAE and MSE when compared to the previous model.

A Boosting Model to Predict Material Properties

As a Postdoctoral Researcher, I spearheaded a team of Ph.D. students to meticulously extract Alloy preparation properties, followed by the creation of over 3000 features derived from Quantum Simulations. Subsequently, I employed multiple Boosting models to forecast four distinct temperatures crucial to understanding a significant alloy property.

The standout performer among these models was a CatBoost model, which not only achieved an impressive average R2 score of 94% in predicting these properties but also demonstrated its ability to accurately predict the four target temperatures for alloys containing elements not present in the training dataset.

Furthermore, I harnessed the power of feature importance analysis, utilizing the Shapley value method to construct a revealing graph. This graph provided invaluable insights into the Quantum properties with the greatest impact on the four target temperatures.

Salt Lake County Housing Insights: Market and Homeowner-Sentiment Analysis

Nerd Realtor In this project, I conducted data scraping to gather demographic information and housing prices for Salt Lake County, organized by zip code. Furthermore, I extracted more than two hundred homeowner postings for each zip code from a neighborhood website.

Employing TextBlob and NLTK, I performed sentiment analysis on these posts, ultimately generating a happiness index for each zip code. To visualize this data, I leveraged Folium's heatmap representation within a web application developed using Streamlit.

Users can select specific demographic, market, and happiness indices, and the resulting overall index is presented in the form of a Folium heatmap. This approach offers an efficient means of conveying a substantial amount of information at a glance.