Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Self-Training via Second-Level Support
This talk demonstrates a self-training AI agent that learns from human replies by extracting and integrating useful answers into its training data automatically.
In this 5-minute demo, I’ll show how I built a self-training mechanism for an AI Agent by leveraging second-level human replies. When the AI fails to answer, the conversation is routed to a human. If the human provides a useful answer, it’s automatically extracted and stored as new training data — allowing the agent to improve its performance over time without manual labeling. I’ll walk through the architecture, data flow, and key functions that make this loop work.
I’ll walk through the key components:
- A webhook that listens to human responses and filters suitable training samples
- A function that reformulates user+human messages into canonical Q&A entries using LLMs
- A script that updates the RAG index (e.g., Pinecone, Qdrant, or local FAISS)
- An optional visual flow using Tiledesk Design Studio to orchestrate the fallback and learning path
Customizable Generative AI Agents automate workflows using a no-code Visual Builder.
Open-source LLM agent builder automating multichannel workflows via Node.js/Kubernetes orchestration.