# Amitech

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Vikramaditya Pratap Singh](https://cerebralvalley.ai/u/vikramaditya)
- **GitHub:** https://github.com/d02ev/arxplain
- **Demo video:** https://www.loom.com/share/39c9bdbaef184c4ea06a0a9eba6a3889
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1068

ArXplain: Research papers, explained at your level

It reads a research paper, finds out what it assumes you already know, asks what you actually know and then explains every section of the paper at exactly your level.

The problem: Research papers are written for experts. A student reading "Attention Is All You Need" hits scaled dot-product attention, BLEU scores, ablations, and unstated assumptions all at once with no way to know which gaps matter or where to fill them. Generic AI summaries don't help: they flatten every paper to the same depth for every reader.

What ArXplain does: It's a CLI-based multi-step agentic system with two phases:

1. Profiling: An LLM, assisted by web search (Tavily/DDGS), analyses the paper and classifies its topics as Hard Requirements or Good To Know. For each topic, you're shown a passage from the paper itself plus a plain-language self-check question, and you rate your familiarity from 1–10. Your answers are saved as a per-paper profile.json.
2. Explanation: ArXplain explains every section of the actual paper: Abstract, Background, Experiments, Results, Conclusion, References in the paper's own order, adapted to your ratings: slowed down with basics where you rated low, brief recap where you rated high. Each section lists its hard requirements; figures appear in the sections that discuss them, references link to the actual cited papers, and every external learning resource lives exactly once in a Resources sidebar that section topics jump-link to.

Why it's different: The explanation is personalised per topic, not per user stereotype; it explains the real paper instead of summarising it; and the LLM pipeline is cost-engineered with one batched call per explanation, hard call caps, and disk caching, so a full paper costs roughly 1,700 tokens.

Built with: Python, Meta Model API (Muse Spark), Tavily/DDGS search, arXiv full-text ingestion with a canonical versioned JSON paper schema.

## More from Meta Global AI Developer Hackathon

- [SunSeg](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1065)
- [Anytime Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1066)
- [Adiyan-Wayfinder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1067)
- [VariantCut](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1069)
- [BlackboxAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1070)
- [Manual for Anything](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1071)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1068. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt