- PROBLEM
- An e-learning platform needed authoring help from a language model that users could trust: suggestions arriving live, in the document, with a way to tell a good answer from a plausible one.
- BUILT
- A course-authoring agent with real-time edits over WebSocket, a multi-provider SDK across OpenAI, Mistral, AWS Bedrock and ElevenLabs, prompt management, phased evaluations judged by models, and guardrails that correct their own output. Alongside it, R&D on hybrid retrieval with source-aware routing and conflict detection.
- SINCE THEN
- Part of 500+ merged pull requests across 70+ repositories on that engagement. The client is not named.