Society of Scientists
1st place at the AGI House × Sundai AI Agents Hackathon at MIT. 9 specialized AI agents on Jamba long-context models that read literature, brainstorm, and draft a grant proposal from a single research prompt.
The project was inspired by Markus J. Buehler and Alireza Ghafarollahi's groundbreaking work on LLMs and AI agents for scientific discovery — see also SciAgents. We extended that direction with agency-specific templates, compliance checking, and a structured group-chat orchestration built on AG2 (AutoGen), with literature search powered by Exa on top of AI21 Labs' Jamba models.
Built in September 2024 at the AGI House × Sundai hackathon at MIT. This was my first hackathon win! Augmenting the scientific process is a topic that I care about deeply - it made the day a special one.
Pitch deck
How it works
You give it a research topic. The system runs a structured multi-agent conversation:
- Domain scientists contribute expertise across their fields
- A researcher integrates literature search (PubMed, arXiv, CrossRef, Semantic Scholar) and manages citations
- A planner synthesizes those inputs into a proposal outline
- Grant writers expand each section — hypothesis, methodology, objectives, budget, ethics, novelty
- A compliance checker validates against agency requirements (NIH, NSF, DOE)
- A quality assessor scores the draft with NIH/NSF-style review criteria
- A critic reads the full proposal and proposes revisions
The output exports to PDF, DOCX, LaTeX, or Markdown with citations in APA, Vancouver, or BibTeX.
The team
Endless thanks to my amazing teammates for making this happen:
And to the judges — Tuan Ho, Mark Weber, and Yaniv Markovski — for the trust and the feedback.