🧑‍💻Updates log

9/5/25

Model:

• Switched to Gemini 2.5 Pro.

Training:

• Enabled RAG with a vector database to incorporate both pre-training data (curated videos, old tweets, writings, etc.) and new knowledge added to the knowledge graph.

Knowledge Base Scraper:

• Now goes beyond summarizing websites, also diving into the three most relevant links to gather detailed information. NOTE: A summary will be displayed in the UI, but the whole content will be stored in the Pinecone vector database.

• Automatically refreshes (by re-scraping) websites in the knowledge base.

Builder Helper (Grant Action):

• Implements a conversational flow to collect the project’s website, documentation, repository, and demo for further evaluation, while considering the user’s profile.

To Improve:

• Autofill the knowledge base with new tweets from Jesse Pollak.

• Improve understanding of dates and numbers in scraped knowledge.

• Improve scrapping service to better capture data.

• Include data gathered from builders’ projects in the “Grants” section of the UI.

• Validate projects using the builder’s score from Talent Protocol.