🎓 Harvard University · Computer Science
Hi, I'm Nathan.
I build ML systems and dig into messy data, treating models like philosophical puzzles to see how human belief maps onto machine logic. I like to understand why models really work.
🎓 Harvard University · Computer Science
I build ML systems and dig into messy data, treating models like philosophical puzzles to see how human belief maps onto machine logic. I like to understand why models really work.
Currently
Experience
Built an on-premise log anomaly detection pipeline using LLaMA-3 (8B), fine-tuned with LoRA on SSD telemetry logs to identify failure patterns at 99% accuracy.
Forecasted ERCOT electricity prices using LSTMs, MLPs, and fine-tuned TimeGPT. Designed a hybrid framework layering a spike-correction MLP over TimeGPT, reducing MAE by 43% and August spike error by 50%.
Led a 6-person team building predictive models for used car pricing. Also collaborated with the Chan Zuckerberg Initiative on housing affordability, building 21 county-level random forest models and improving R² by ~15%.
First-authored a paper on predicting suicidal tendencies in tweets using LSTM and Transformer models, published in IEEE and presented at the 4th Asia Conference on Information Engineering.
Conducted meta-analysis reviewing 258 abstracts on predictors of social media popularity. Identified 53 articles for full-text review and extracted key performance metrics.
Projects & Research
Post-training quantization framework for LLMs. Extended SpinQuant with learned diagonal affine transformations to fight weight outliers, cutting reconstruction error by up to 60% on anisotropic datasets and outperforming 4-bit baselines on GPT-2.
A political simulation site that models polling preferences and compares election systems: popular vote, ranked-choice, and ranked-pair. Built it because I wanted to see what would actually happen.
Contact
Open to research roles, internships, and conversations about interesting problems.