Shabbos Prep Agent

Inspiration

Every Friday, millions of observant Jewish households race a hard deadline that moves every week and differs by city: candle lighting. After it, cooking, driving, phones and lights are off for 25 hours, so anything not done is simply not done. In my own community (Lakewood, NJ) Friday is a logistics problem: guests, a full menu, kids, a work day, and a sunset that comes at 4:15 in December. Existing apps tell you when Shabbos starts. Nobody tells you what to do, and when, to be ready.

What it does

You describe your Friday in plain English ("8 guests, I finish work at 1, soup, roast, kugel, two kids need baths, we light 18 minutes before sunset"). The agent:

1. Calls real tools for the real times (candle lighting, havdalah, parsha, Yom Tov, zmanim) for your city.
2. Back-schedules every task from candle lighting with buffers (oven times, the urn and blech, timers, baths, table, a final 15-minute "everything off" margin).
3. Returns a time-stamped plan, a short message to forward to the family group chat, and an .ics file so every phone pings 10 minutes before each step.
4. Shows its work: an agent trace of every tool call it made. It never guesses a time and never gives halachic rulings ("ask your Rav").

How we built it

• Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) on Nebius Token Factory is the planner: an OpenAI-compatible tool-calling loop (get_shabbos_times, get_zmanim) followed by a strict JSON plan.
• Nemotron 3 Nano (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B) writes the family message: the fast, cheap everyday call, exactly the Super-for-reasoning / Nano-for-speed split the track recommends.
• Model IDs are env-configurable (e.g. nvidia/Nemotron-3-Ultra-550b-a55b as planner).
• Hebcal.com (free, CC BY 4.0) supplies the times; a cached client keeps us polite.
• Gradio UI hosted on Render; .ics export with alarms; pytest tests with a scripted model so the agent loop is tested offline.

Challenges

• The model doesn't know today's date. In our first live run it confidently planned a Friday in 2025. Fix: we inject today's date and the coming Friday into the prompt, then a grounding step overwrites candle lighting, havdalah and parsha with the tool's real values and warns about any task that is on the wrong day or eats into the 15-minute buffer.
• Models still sometimes squeeze a task into the last 15 minutes. A self-check step now catches it, hands the violations back to Nemotron 3 Super, and shows the corrected plan.
• The summary model once invented a candle time; it now gets the verified time explicitly. Nano is a reasoning model, so a tight token cap could leave it with no answer; we give it room and keep a deterministic fallback message.
• Getting strict JSON back reliably after tool calls: a tight output contract plus a tolerant extractor.
• Time zones and DST (this hackathon ends the week clocks change): all times are ISO-8601 with offsets.
• Being useful without overstepping: the agent plans logistics but defers halachic questions.

Accomplishments

• A real, end-to-end agent a family can use this Friday, with times that match Hebcal to the minute.
• Two-model routing: a full plan plus family message comes back in about 15-25 seconds.

What we learned

Nemotron 3 Super is a strong tool-caller for structured planning; Nano is more than enough for short, warm prose. Token Factory's OpenAI compatibility meant zero SDK changes.

What's next

WhatsApp delivery and a Friday-morning check-in ("you're 20 minutes behind — move the kugel up"), saved household profiles, Yom Tov meal-by-meal planning, and Hebrew/Yiddish output.

Feedback on Nebius Token Factory and NVIDIA

Signup and API key creation were quick. The OpenAI-compatible endpoint worked with the stock openai Python SDK and Nemotron 3 Super handled tool calling well. The main issue was that the model does not know today's date, so we ground it with the real date and tool results. Model ID casing differs between models (nemotron-3-super vs NVIDIA-Nemotron-3-Nano), which is easy to get wrong; consistent IDs or aliases would help.

Is this an existing project?

No. It is a new repository written during the submission period. The author's earlier, separate project (shabbos-zmanim-mcp, an MCP server for voice assistants) shares the domain only; no code was reused.

Built With

• gradio
• hebcal
• icalendar
• nebius-token-factory
• nvidia-nemotron
• openai
• python
• render