Anti Patterns

Common failures that waste time and money. Extracted from AI Engineering Anti-patterns by Chip Huyen

⛔️ Using GenAI Unnecessarily

Teams adopt generative AI for status, not solutions. Companies want bragging rights: "we use AI."

Example: Energy startup built GenAI for optimisation. A simple algorithm checking cheap hours works better.

⛔️ Premature Complexity

Teams rush to complex solutions first. "I must climb that mountain" mentality. Some common mistakes

  • Vector databases before basics

  • Agent frameworks too early

  • Fine-tuning before prompting

⛔️ Bad Product or Bad AI? or Both?

💡 Technical excellence ≠ product success.

Meeting summaries example:

  • ❌ Summary length (3 vs 5 sentences)

  • ✅ Action items from meetings

LinkedIn job recommendations example:

  • ❌ Match accuracy scores

  • ✅ Skill gap identification

Tax chatbot example:

  • ❌ Users don't know what to ask

  • ✅ Suggest relevant FAQs

⛔️ Not Understanding Users

User experience means understanding human behaviour.

💡 Users want help, not judgment.

⛔️ Overcomplicating RAG Systems

Teams rush to vector databases.

  • Start here: Keyword search, clean data, smart chunking

  • Data prep matters: Add context, include titles and keywords

⛔️ Using Agent Frameworks too early

Agent frameworks create hidden complexity.

  • ❌ Framework bugs become your bugs

  • ❌ Starter prompts contain typos

  • ❌ Performance fluctuates with framework changes

  • ❌ Hard to diagnose root causes

Direct API calls for simple tasks beat frameworks.

  1. Build core logic with direct API calls

  2. Measure performance with 100+ test cases

  3. Identify specific bottlenecks frameworks solve

  4. Only then evaluate if framework complexity is worth it

⛔️ Lack of Human Evaluation

AI judges help but have limits. Add human evaluation.

Limitations:

  • Models change

  • Hard to track progress

  • Teams lose touch with users

  • Tools score differently

⛔️ Not Localising Failures

Multi-step AI needs failure isolation.

  • ❌ Resume extraction fails. Which step? PDF to text? Text to data?

  • ✅ Debug each step separately.

⛔️ Overindexing on Early Success

Weekend demos don't predict launch timelines. Demo = 20% of total work.

Production needs:

  • Error handling

  • User testing

  • Performance tuning

  • Infrastructure

  • Security

  • Monitoring

  • ...and more

⛔️ Giving Up Too Early

Teams quit after first failures. Find root causes. Fix them. Common causes:

  • Bad product design

  • No user research

  • Poor evaluation

🤓Issues

Last updated