Newest first.
MSE capstone for pediatric speech-language clinics. LLMs extract clinical signals from therapy notes, one numeric record per client, with leakage windows so each feature only sees allowed sessions. YAML signal registry, Streamlit ops UI, FastAPI + Next.js clinic app, Compute Canada batch runs, LLM judge validation, Supabase publishing.
Ingests SEC EDGAR financial statement data into BigQuery with Python. dbt models for staging, intermediate, and mart layers, with schema tests and lineage docs across companies and fiscal quarters.
Pulls articles from the Guardian API, chunks and embeds them in Supabase pgvector, and uses retrieval to ground Gemini briefings. Orchestrated in YAML with incremental ingestion, Parquet intermediates, and GitHub Actions CI on every push.
SYDE 577 final project: decode neural signals into text. HDF5 dataloader for brain recordings, swappable LSTM/GRU/RNN/Transformer encoders, CTC loss on phoneme sequences.
Built autodiff from scratch in NumPy for SYDE 577. Parameter, Input, and op nodes in a computational graph with forward and backward passes, matrix multiply, activations, and a trainer on pickled test parameters.