I built an internal workbench for editorial teams experimenting with AI generated content. The goal: shorten iteration loops without losing tone or accuracy.
The assistant orchestrates prompt templates, fact checks with retrieval, and delivers drafts in structured blocks editors can remix.
What I built
- Designed streaming UI for token-by-token drafts with editor controls.
- Implemented retrieval augmented generation pipeline with pgvector.
- Built experiment tracking to compare prompt and model performance.
Impact
- Editors cut draft turnaround time by 55% with structured workflows.
- Prompt eval dashboard highlighted best-performing prompt/setting combos.
- Safeguards reduced hallucinations by gating responses through fact checks.
