**This is a repost - Post was deleted by mistake

https://lyon-craft.fr/sessions/l-ia-dans-vos-projets-un-peu-beaucoup-pas-du-tout.html

Speaker: Marie-Alice BLETE – 10+ years as Java dev, now specializing in AI –Co author of Developing Apps with ChatGTP publish @ O’Reilly – Currently working at Komodo Health

1. Live AI-Powered Web Interfaces

  • Demo of a website shop with AI integration.
  • Users can request new features live, which are then added via AI in real time.
  • Demonstrates the potential of LLMs to drive interactive, on-demand UX.

2. AI Agents: Autonomy & Challenges

  • Agents operate based on objectives (e.g., via LangChain framework):
    • Prompt defines goals.
    • Agents are given access to tools (e.g., Python function for product search).
    • Each tool must be explicitly defined as usable.
  • Pros: Can act autonomously.
  • Cons:
    • Hallucination risks (like any LLM).
    • High cost, especially for simple use cases or POCs.
    • Unstable behavior due to the evolving nature of LLM providers.

3. Architecture Considerations

  • Traditional: Monolithic “black-box” agents.

    • Difficult to maintain and scale.
    • Adding new capabilities may disrupt existing ones.
    • Requires larger context → higher costs.
  • Alternative: Modular multi-agent systems.

    • Each agent has a specific task.
    • Allows tracing and debugging.
    • Downside: Rigid scenarios → poor generalization outside designed use cases.

4. Development Lessons (Maplab Internal Use Case)

  • Example: Drug search system avoids LLM for final output → Uses LLM to generate API queries (like text-to-SQL).
  • Production Learnings after 9 months:
    • Team composition: Mix of devs, ML engineers, researchers. Slower but more comprehensive.
    • Estimations: LLM work is less predictable than traditional dev.
    • Testing: Shift from unit testing to system-wide evaluations. Harder to guarantee output quality.
    • Tech stack: Constant reevaluation due to rapid evolution of models, providers, capabilities.
    • RTE is unreliable due to novelty of the domain and lack of community experience.

5. Q&A Highlights

  • Q: How to prevent hallucinations?
    A: No silver bullet. LLMs are designed for completion, not Q&A. RAG (Retrieval-Augmented Generation) helps, but building a reliable system takes time. Plug-and-play doesn’t work.

  • Q: Can LLMs be told to stay silent if unsure?
    A: Technically no. LLMs don’t “know” what they don’t know. Tools like Inspector Agents can flag hallucinations, but they double the cost.

  • Q: ROI of implemented features?
    A: Still early. Clients appreciate the chatbot UX (e.g., spreadsheet-powered search), but hard to quantify return. It is a strong marketing differentiator.

  • Q: How is data security ensured (e.g., for medical data)?
    A: LLMs don’t access data directly. They generate queries sent to secure APIs. The system is designed so that LLMs only act as interpreters, not data processors.