Lukas Gearin

Engineered a predictive financial model implementing a stacked LSTM architecture with GRU units to predict next day stock prices by leveraging 5 day moving averages. Model achieved a MSE of 5.36e-5 with trend accuracy of 82%.

Building an NLP system using FinBERT, spaCy, and transformers to detect financial catalysts in news, integrating rule-based and ML driven sector mapping with an XGBoost directional impact classifier to forecast sector movements with confidence scoring.

Implemented a transformer-based text-to-speech system leveraging VCTK data, incorporating BERT-based text tokenization, speaker conditioned encoding, LSTM-based mel spectrogram decoding, and a Griffin-Lim vocoder to generate synthetic speech.

Developed a character-level transformer-based language model to generate original text in the style of any inputted works, implementing multi-head self-attention, positional embeddings, and autoregressive decoding for creative language synthesis.

Built a Retrieval Augmented Generation (RAG) system for bankruptcy document analysis, integrating sentence chunking, semantic embedding with SentenceTransformers, and LLM-based answer generation using Hugging Face models for precise financial insights extraction.

Designed and implemented a Django web application enabling encrypted messaging, where encryption keys are algorithmically derived from pixel matrices of user uploaded images, introducing image-based entropy into the encryption process.

Implemented a NBA prediction system leveraging gradient boosted random forests paired with deep neural networks yielding results of 91% accuracy. Combines traditional ML with deep learning for superior sports analytics.