Practical tips for building AI applications using LLMs
summary
Kushu AI builds LLM apps and reports practical fixes: expect 1–5% error rates (versus ~0.1% for stable APIs), add retries/timeouts from day one using Tenacity (Python) or npm retry and LangChain’s retry output parser, validate structured outputs with Pydantic schema validation or Yup and re-prompt on failure, prefer Server‑Sent Events for simple streaming and Cron → RQ task queues for minutes‑long jobs, run prompt unit tests with Promu evals, and benchmark OpenAI (GPT‑4/4o), Claude for code, Gro for latency, Bedrock for data residency.
AI-written summary · about
chapters (9)
- 0:00Introduction to AI Application Development
- 1:02Handling LLM Inconsistencies
- 5:32Implementing Retries and Timeouts
- 10:33Managing Structured Outputs
- 15:13Implementing Streaming in LLM Applications
- 20:53Using Background Jobs for Long Tasks
- 29:09Evaluating Prompts with Evals
- 31:53Overview of Foundational Models
- 38:03Conclusion and Contact Information
related talks
15mPractical tactics to build reliable AI apps
43mAI Development Lifecycle: Learnings of What Changed with LLMs
27mBuilding AI applications: Successes and Lessons
47mEmbedding AI: Don’t Burn Your Fingers in the Heat of the AI revolution
36mThe Best AI Strategy? Don’t Build AI 🤯 Build AI-Ready Software
30mAI Application Architecture: Building for Scale and Efficiency
46mEnhancing Agent Reliability Through Evaluation-Driven Development
24mBuilding Real-Time AI Decision Systems with LLMs