LLM-Based Development
summary
A prototype booking app wires voice transcription to GPT-4 and GPT-3.5 as the decision engine, calling a calendar API to schedule appointments; one demo successfully books “Dr Smith, Wednesday 10:30,” while another shows GPT-3.5 returning inconsistent internal state and hallucinating availability. Architecture uses system/controller/user roles and forces temperature=0 for determinism, with the model emitting discrete action tokens consumed by the backend. A lightweight testing framework for LLM-based development is under development and planned for open source to catch regressions.
AI-written summary · about
chapters (9)
- 0:00LLM-based development intro
- 4:21What LLMs can do
- 7:18LLM appointment system architecture
- 14:19Live demo: GPT-4 scheduling
- 16:41Live demo: GPT-3.5 scheduling
- 22:26Agent-based development
- 23:33Testing LLM-based apps
- 24:35How testing works
- 27:39Future directions
related talks
19mTesting LLM-Powered Applications
46mBreaching LLM-Powered Apps: Overcoming Security and Privacy Challenges
01h 15m10x Development: LLMs For the working Programmer
24mLLMs: Powering Lightning Fast Product Innovation
01h 00mFrom LLM Theory to Practical Agentic Implementations
35mBuilding LLM-based Agents
01h 59mLLM applications with LangChain
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges