Built, launched, and still operating solo — 10,000+ articles published with no human in the loop. A Python service polls seven wire and national news feeds every 30 minutes, embeds each article, and applies recency-weighted hierarchical clustering to find cross-source clusters of two to five independent reports on the same event. GPT-5 synthesizes each cluster into an original article in a single structured-output call; a fact-grounding pass then verifies the draft against its source texts and rewrites or discards anything it can't support, and embedding similarity against the published corpus blocks near-duplicates. Posts ship via the WordPress REST API with generated headline, categories, source attribution, and bidirectional related-story links. Reasoning-model spend is confined to writing and fact checking; classification and fallbacks run on a cheaper model.

Junk-fee and drip-pricing risk screening: a free assessment scores a business's pricing and disclosure practices, and a paid tier generates a detailed PDF report. Next.js and TypeScript on Prisma/PostgreSQL, Stripe Checkout with webhook-driven fulfillment, React PDF generation, and Amazon SES delivery, deployed in Docker on self-managed CapRover infrastructure. Shipped end to end in a weekend; roadmap driven by competitive analysis.

Working prototypes that answer questions over a university unit's own published content and documentation, rather than sending people to a search box: document ingestion and chunking, embedding and vector retrieval, and grounded generation wired to a chat interface. Paired with reusable AI automation frameworks for repetitive internal editorial and support work. First serious LLM work inside the group.