Machine learning & data science
December 2025Won the $10,000 Unsloth Prize · Google Gemma 3n Impact Challenge
Dream Assistant
A voice assistant that understands atypical speech, and never sends audio off the phone.
ResultAfter fine-tuning Gemma 3n with Unsloth on the cleaned, augmented audio, the model reached a 6.3% word error rate on atypical speech.
Problem & approach
- Problem
- Voice assistants fail people with speech impairments, and sending their recordings to the cloud is a privacy risk.
- Approach
- Collected about 200 recordings of everyday phrases from a person with a speech impairment, then trimmed silence, normalized volume and augmented each clip with noise and speed changes to make the model robust. Fine-tuned Google Gemma 3n with Unsloth so it runs on the phone, and built the Android app and its UI end to end.
Gemma 3nUnslothAudio preprocessingPythonAndroid
2025Top project · Erdős Institute
Predicting drug potency with graph neural networks
Can a neural network read a molecule’s structure and predict how potent it will be as a drug?
ResultThe graph networks reached 70–74% classification accuracy, while a classical QSAR baseline reached 83%: a clear finding about when deep graph models are, and are not, worth it.
Problem & approach
- Problem
- Testing drug candidates in the lab is slow and expensive. Predicting potency from structure narrows the search before anything is synthesized.
- Approach
- Our team of five turned 529 TRPM8 compounds from ChEMBL into molecular graphs with RDKit and compared five graph network architectures for classification and regression. I built and tuned the graph convolutional network (GCN).
GCNRDKitPython
2024Erdős Institute
Predicting loan default
Who will default on a loan? A model trained on 1.5 million real applications that has to stay reliable over time.
ResultThe best model, LightGBM, reached an AUC of 0.778 on held-out test data (0.782 validation, 0.792 training), so it generalizes with little overfitting.
Problem & approach
- Problem
- Lenders need to estimate default risk before approving a loan, and a model is only useful if it keeps working as their clients change. Built on Kaggle’s Home Credit – Credit Risk Model Stability challenge.
- Approach
- Working with one teammate, I cleaned and encoded Home Credit Group’s dataset, handled missing values, and engineered weighted days-past-due features that improved every model. Compared LightGBM, XGBoost and multilayer perceptrons.
LightGBMXGBoostMLPPython
2023Project with distinction · Erdős Institute
Meow by Meow
What is your cat trying to tell you?
ResultAugmentation significantly improved both models, which each reached over 90% accuracy.
Problem & approach
- Problem
- Classify a cat’s meow as comfortable, uncomfortable or hungry from a short audio recording.
- Approach
- In a team of five, we converted recordings to mel spectrograms. I proposed and implemented the data augmentation (time shifts, stretching, frequency masking) and built the convolutional network with a teammate, alongside a k-nearest-neighbours model.
CNNk-NNAudioPython
Side projects
I love building in my free time, and I vibe coded these products.
LiveKarmaKit
A place where builders get their first users and real feedback on new ideas. People swipe through projects, leave feedback, and earn karma points for helping.
karmakitapp.com ↗Web appCommunity
LiveAgentMarket
A marketplace where AI agents hire each other: agents take on projects, subcontract other agents, and sell the strategies that worked to the rest of the market.
agentmarket.online ↗AI agentsMarketplace
Entrepreneurship
Co-founder & CEOCycleBuddy
A women’s health app for tracking your cycle, learning about hormone health, and planning around your own rhythms, with your data encrypted and under your control. It also offers a stigma-free curriculum for schools.
cyclebuddy.ca ↗ProductPrivacyHealth education