Granite Tait

Built in 48 hours for the Utah Tech Week Hackathon where it became a finalist. The system uses Gemini 2.0 Flash to instantiate 4 board member personas with distinct perspectives and debate styles. Each agent reads context from the founder's Google Drive before weighing in. The real innovation is the debate loop — agents argue until they reach a consensus, which is summarized into an actionable recommendation. Key challenge: preventing agents from simply agreeing with each other required explicit dissent prompting in the system message.

An open-source multi-agent morning planning system. Runs as a Python DAG (Directed Acyclic Graph) overnight. Five agents run in parallel where independent: WeatherAgent, CalendarAgent (Google Calendar API), FitnessAgent (Strava API), PriorityAgent (reads a markdown file). SynthesisAgent waits for all and produces a formatted daily brief. Key insight: constraining upstream agents to output structured data (not prose) made the synthesis dramatically cleaner. Delivered via email each morning.

Internal tool built at Paycor to address a real pain point: when bugs arrive, figuring out who on a 100+ person team has the relevant context was slow and inconsistent. The system vectorizes 9 years of Jira ticket history using OpenAI embeddings. New bugs are summarized by GPT-4o, then semantically matched against historical bugs. The top-matching developer is auto-tagged in MS Teams with the relevant historical context attached. Reduced triage time from hours to minutes.

Built for construction field teams who still use napkin sketches to document change orders on job sites. Uses OpenAI Vision to interpret sketches, extract dimensions, and estimate materials. Outputs a professional SVG-based change order document with digital signature support for on-site approval. The key challenge was getting reliable dimension extraction from rough hand drawings — solved with a structured prompt that asks the model to output JSON before generating the document.

A full RAG pipeline: documents are chunked, embedded with OpenAI's text-embedding-3-small, and stored in Supabase's pgvector extension. Queries are embedded at runtime and matched via cosine similarity. The UI shows which document chunks contributed to the answer. Built as a demonstration of how RAG can replace keyword search in document-heavy workflows.

A Playwright-based test suite integrated into Azure DevOps CI/CD pipelines at Paycor. Covers E2E, integration, and regression scenarios. AI-generated test plans use GPT-4o to analyze PR diffs and suggest which test suites to run and generate new test cases for changed logic. Quality guardrails prevent merges when AI-stochastic output variance exceeds a threshold.