Bridging the gap between industry and academia by giving local businesses access to AI/deep learning expertise. Each semester we center the course content around a client's business problem, and equip students with the skills and knowledge to meet the challenge.

Developing complex multi-agent LLM systems to address difficult challenges in wealth and asset management.LLM projects:- Prompt learning for better retrieval - Fine tune and bootstrap examples to learn the best prompt for a taskFinancial advisor chatbot - Aided in portfolio creation given a customer's risk tolerance and investment goalsInconsistency detection in earnings calls - Multiple agents that extract and examine different portions of an earnings callSummarize video and PDF presentations into marketing memo for investors - Transcribe video and extract info from PDFs, summarize, and then combine the information from both into a professional investor memoSentiment for stock direction prediction - Aggregate multiple news articles for a given ticker - Extract novel portions of news - Compare news to price movement after news releaseCompliance adherence system - Multiple agents that validate, summarize, and compare new documents (e.g. SEC proposed rule, marketing materials) to SEC regulations to ensure compliance - Combination of GraphRAG and RAG databases - Fine tuning RAG embeddings to reduce hallucinations and improve response accuracy - Distillation from closed source models for better capabilities - Fine tuning LLM instruction learning for specialization

Creating deep learning algorithms for client software.- Overseeing all quantitative research including discrete choice models, customer segmentations, and key driver analysis.- Developing interactive simulators to make the output of discrete choice models useful to executives.

Designing and building equity trading models for systematic trading. This includes:- Guiding a small team in the development of deep learning models for time series forecasting, tabular data modeling, self-supervised pretraining, hyperparameter optimization, reinforcement learning for portfolio optimization, and back testing methods.- Implementing multimodal models using Large Language Models (LLMs) for text, time series and tabular data.- Building specific trading constraints provided by financial analysts into the model's learning.- Designed and implemented proprietary consistency learning method to enhance classification.- Participated on sales calls as technical expert.- Appointed to Board of Trustees, and helped make pivotal changes to the business.

Developing text augmentation and semi-supervised deep learning models for data labeling using LLMs and consistency learning techniques.

Developed novel multi-modal attention model that combines customer behavior and characteristics to predict future purchases. Augmented NLP classification models with external datasets and attention mechanisms.

Currently, working on image segmentation, facial recognition, and neural machine translation.

I develop deep learning solutions and code to a variety of problems:- Project Poseidon: Applying software-defined networking and deep learning to security. Uses hierarchical recurrent networks to encode sessions. With a trick of manually creating anomalous traffic, 84% of attacks were caught with a 0.5% false positive rate on the ISCX IDS dataset. The associated repo was awarded rookie open-source project of the year for network security (https://www.blackducksoftware.com/open-source-rookies-2016).- Project D*Script: Author identification in handwritten documents. Used a combination of convolutional autoencoders and manual feature extractors to make handwritten documents searchable. These features were used to search a database for similar styles of handwriting. This architecture achieved 3x higher accuracy over the best commercial solution. http://ieeexplore.ieee.org/document/7926719/. - Project Sunnyside Up: Sentiment analysis. Used a character-level CNN to classify sentiment in text. It achieved state of the art results on the IMDB movie review dataset. You can find my blog posts on gab41.lab41.org and blog.cyberreboot.org.

Helped create and teach a new course on quantitative business research.- Solicited local companies as research partners each semester.- Students collected and analyzed real world data related to the customers of the sponsoring companies. At the end of the semester the findings were presented to the executives.- Quid pro quo with sponsoring companies: expensive research at no cost in exchange for letters of recommendation and internships for the students.We took the Utah State Fair from its worst revenue year in history to their best revenue year.