Reflections on local-first LLM architectures, quantitative forecasting, machine learning pipelines, and software design.
Why local-first architecture matters for technical learning. How we combined Markdown files as the single source of truth, ChromaDB for semantic vector embeddings, and Ollama for zero-cost private inference.
An exploration into assembling multi-horizon predictors. Extracting sentiment polarity from financial news feeds with FinBERT and feeding engineered rolling indicators into gradient-boosted ensembles.
Separating presentation layers from core backend services. How routing lab, blog, and portal across distinct GitHub repos and Cloudflare edge deployments eliminates single points of failure.