> For the complete documentation index, see [llms.txt](https://a00-4.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://a00-4.gitbook.io/docs/ai-agent-tech.md).

# AI-Agent Tech

The agents training architecture is built to operationalize onchain leaders knowledge, judgment, and public presence into an intelligent, always-on agent. It is designed to scale support to thousands of builders with practical advice, funding intelligence and other onchain capabilities. The system combines fine-tuned language modeling with Retrieval-Augmented Generation (RAG), active learning loops, and real-time integrations.

<figure><img src="/files/jDhjgzbMIADCkUVbp2CU" alt=""><figcaption></figcaption></figure>

### 🎯 Training Goals

* **Capture Original's Persona:** Reflect leader's tone, decision-making, and domain fluency.
* **Stay Fresh & Real-Time:** Sync continuously with builder queries and ecosystem updates.
* **Learn from Feedback:** Incorporate the original's feedback and user signals in daily model updates.
* **Support at Scale:** Maintain high-quality interactions across Farcaster, X, Telegram, TBA, XMTP and more.

***

### 🧠 Training Pipeline Components

#### Pre-Training: Building Onchain's Minds Persona

* **Goal:** Establish baseline persona and communication style.
* **Sources:**
  * YouTube videos & podcasts
  * Historical posts on X
  * Farcaster threads and replies
* **Curation Process:**
  * Transcription → Cleaning → Synthetic Sample Generation (via Gemini)
* **Output:** Base model aligned with the leader's tone and expertise.

#### Fine-Tuning: Specialization & Personality Alignment

* **Model:** Gemini 2.5
* **Data:** Curated public content + dashboard personalization (bio, tone examples, answer style)
* **Focus Areas:**
  * Align tone's communication
  * Minimize hallucination or generic output
  * Embed optimism, builder-first mindset

#### RAG System: Real-Time Contextual Intelligence

* **Vector DB (Pinecone):**
  * Structured by namespace
* **Ingested Sources:**
  * Websites
  * Farcaster + Twitter posts (real-time)
  * PDFs, notes, URLs, GitHub (via Puppeteer w/ refresh)
* **Latency Optimization:**
  * Caching, response reranking, fast retrieval
* **Moderation:**
  * Filters for PII, toxicity, and spam
* **Pending:** Intent recognition, data governance, sentiment engine, Knowledge Graph enrichment

#### Feedback Loop: Active Learning + Evaluation

* **Human-in-the-loop:** Active reviews and scores responses in the Agent Dashboard
* **Pipeline:**
  * Good responses → Reinforced in training
  * Bad responses → Flagged and retrained
* **Live Model Updates:** Responses are iteratively polished and personalized via ZEP layer:
  * Dialogue tracking, intent classification, profile-based refinement

With a base personality, active learning from dynamic and static sources, plus the human-in-the-loop feedback, an agent can improve over time to be the best companion for builders in any ecosystem.