💡Scaling Builder Support with a Digital Sidekick

JesseXBT Agent Training Documentation (v2)

🔌 Agent Training Overview

The training architecture for jessexbt is built to operationalize Jesse Pollak’s expertise, style, and values into an AI-native agent capable of supporting 1000+ builders daily. This pipeline integrates fine-tuned language modeling, Retrieval-Augmented Generation (RAG), and feedback-driven improvement loops.

⚡️ Current Phase: Phase 0 - MVP

🌟 Training Objectives

• Embody Jesse's Persona: Reflect Jesse's tone, knowledge, and judgment.

• Stay Real-Time: Leverage RAG to remain up-to-date via active ingestion.

• Iterate Daily: Use Jesse’s ongoing feedback for continual improvement.

• Scale Builder Support: Enable scalable, personalized guidance across Farcaster, X, and Telegram.

---

🧠 Fine-Tuning Pipeline

Pre-Training: Persona Construction

• Sources:

  • 164+ videos & podcast appearances (transcribed)

  • X (Twitter) posts

  • Farcaster interactions

• Processing:

  • Transcription (via Gemini 2.5)

  • Cleaning and deduplication

  • Synthetic sample generation

Fine-Tuning: Expert Alignment

• Model: Gemini 2.5

• Inputs:

  • Curated media/text dataset

  • Personality settings from Agent Dashboard

  • Hand-authored example responses

• Goal: Minimize deviation from Jesse’s tone; boost alignment on key topics (Base, Web3, funding)

---

📂 RAG System: Real-Time Knowledge Enhancement

Vector Storage

• Technology: Pinecone

• Namespaces: jessexbt, builders, protocol knowledge

Ingestion & Retrieval

• Scraped Data (via Puppeteer):

  • base.org

  • GitHub profiles, demo links, protocol sites

  • PDFs, URLs, notes (manual + auto refresh)

• Live Feeds:

  • Farcaster, X, Telegram (builder queries, Jesse’s posts)

  • Real-time sentiment tracking

Generation

• Retrieved chunks + fine-tuned model → RAG-enhanced response

• Latency Optimization: Caching & fast reranking pipelines

• Moderation Layer: PII, toxicity, abuse filtering

---

🧑‍💻 Feedback Loop: Active Learning

Loop Design

• Human-in-the-loop: Jesse evaluates answers directly

• Evaluation Interface: Agent Dashboard

• Scoring Dimensions: Accuracy, tone, relevance

Update Mechanism

• Positive: Reinforcement into fine-tuning set

• Negative: Flagged for targeted retraining

• Continuous Updates: Ongoing model refresh with new logs + corrections

---

🔗 Technical Integration Overview

Component	Description
Model	Fine-tuned Gemini 2.5 on Jesse’s voice and context
Storage	Pinecone vector DB (multi-namespace)
Feedback	Active dashboard scoring + retraining
Platforms	Farcaster, X, Telegram ingestion + delivery
Moderation	Real-time PII/toxicity filters in RAG system

---

🌌 Why This Pipeline Works

The jessexbt pipeline is crafted to be:

• Personified: Authentically represents Jesse’s tone and style

• Real-time: Constantly updated with fresh inputs and community pulse

• Feedback-driven: Learns from every Jesse interaction

• Scalable: Designed for 24/7 interaction with hundreds of concurrent builders

It’s a digital sidekick with memory, opinion, and Base-native fluency—aimed at removing bottlenecks between builders and actionable help.

---

Last updated: May 8, 2025 Current Phase: MVP (Phase 0) Up Next: Phase 1 — Core Infrastructure and Alignment