Saba Yosef: An AI Agent That Prepares for Shabbat
Connecting Israeli news to the weekly Torah portion — a full-stack agentic system powered by Claude
Every week, Jewish families gather around the Shabbat table and share a Dvar Torah — a short insight connecting the weekly Torah portion to life. Preparing one well takes time: reading the parsha, finding relevant commentaries, and weaving in something timely and meaningful.
Saba Yosef is an AI agent that does this automatically. Every Thursday at 8 AM, it pulls the latest Israeli news, retrieves the weekly parsha with commentaries from Sefaria, and uses Claude to map thematic connections between current events and Torah wisdom. The result is an interactive dashboard where you choose the themes that resonate, and the agent generates five polished Dvar Torah suggestions — ready for Shabbat.
How It Works
The system runs a multi-stage pipeline:
-
News Aggregation — RSS feeds from six Israeli sources (Ynet, Walla, Maariv, Globes, Yisrael Hayom, Google News Israel) are scraped and parsed via feedparser and BeautifulSoup.
-
Torah Retrieval — The current parsha is fetched from Hebcal, then enriched with 60+ commentaries from Sefaria across five categories: Pshat (literal), Hasidic, Ethical, Midrashic, and Biblical Criticism.
-
Connection Mapping — Claude identifies thematic bridges between news stories and Torah topics, producing an interactive SVG connection map.
-
Generation — Based on user-selected themes, the agent generates five Dvar Torah drafts with proper source attribution.
-
Editing & Export — A rich-text editor with AI chat assistance lets you refine the output, then export a formatted PDF with sources.
The Tech Stack
Backend: Python 3.12+, FastAPI, SQLModel/SQLite, WeasyPrint for PDF generation, Claude CLI (no API key required).
Frontend: React 18, TypeScript, Vite, Tailwind CSS, Tiptap editor with full RTL/Hebrew support, Server-Sent Events for real-time streaming.
Data Sources: Sefaria API, Hebcal API, six Israeli news RSS feeds.
Why Build This?
This started as a Friday project — something personal, something fun. But it’s also a case study in what agentic AI looks like when it’s grounded in a real workflow:
- Multi-source orchestration — the agent coordinates across news feeds, Torah APIs, and an LLM, each with different data formats and failure modes.
- Human-in-the-loop design — the connection map lets the user steer the generation, not just consume it.
- Hebrew/RTL handling — a non-trivial engineering challenge that most AI demos skip entirely.
- Autonomous scheduling — the cron job means the system is ready before you are.
Named after my grandfather Yosef, who always had the best Dvar Torah at the table.