# Orchanium

- **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:** [Moayad Moatamed](https://cerebralvalley.ai/u/Moayad)
- **GitHub:** https://github.com/moayadMoatamed/Orchanium
- **Demo video:** https://www.youtube.com/watch?v=VFYUAe-Ta9s
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/578

Orchanium answers a question every drug team faces: we know the molecule we want, but how do we actually make it?

Scientists can design a new molecule that should treat a disease, but finding a way to build it (which buyable chemicals to start from, which reactions in which order, under what conditions) means working backwards through the literature. This is retrosynthesis, and today it takes weeks of reading plus many costly failed attempts in the lab.

Orchanium's strength is not research alone or reasoning alone, but the loop between them:
1. Take it apart: the target is broken into possible steps with real chemistry rules (RDKit), down to buyable chemicals.
2. Research: agents on Meta Muse Spark search the literature live for how similar molecules reacted in similar cases, including conditions, yields and failures.
3. Build a knowledge graph: every piece of evidence is connected to the step it supports and graded by how closely it matches the target's own chemistry.
4. Debate: a Builder agent proposes a route over that graph and a Skeptic agent attacks every weak step. Gaps the debate exposes send the researchers back for more evidence, and the route is revised.
5. Decide: a fixed rule over the evidence, not the model's opinion, computes the verdict, and the output is a step-by-step method with the source behind every step and what to test first.

This even works for molecules that have never been made. Give Orchanium a brand-new molecule today and it finds the building blocks, looks at how those pieces behaved in comparable published reactions, compares them to the new target, and argues out the most defensible path through a network of connected evidence. Instead of starting from a blank page, the chemist starts from the route most likely to work and knows exactly which step is riskiest.

The chemist stays in the loop. After running a step, they log the result as a lab note. It becomes the strongest evidence in the graph, the verdict is recomputed with every change shown, and every later run remembers it. Each experiment makes the next answer better grounded, step by step, until the molecule is made, which means fewer failed attempts, less time and less money per new drug.

To prove it reasons rather than looks things up, blind mode hides every source about the target. On the cancer drug sotorasib, with 41 documents hidden (including the makers' own paper), Orchanium reached the same five building blocks and the same four reactions as Amgen's published manufacturing route. Two steps were in the opposite order, and that is the step Orchanium flagged as unproven. On six held-out patents, every miss came back "Uncertain", never a confident wrong answer.

Everything in the submitted repository was built during the hackathon (a separate repo with a different architecture, not a continuation of the old codebase: no code was carried over, and the commit history starts 30 Oct 2026). It evolved from an earlier proof of concept I made, which was one LLM proposing steps over a hand-loaded database, into a new multi-agent system: live literature research, an evidence graph built per run, a Builder/Skeptic debate, a verdict computed by fixed rules, blind-mode validation, and lab notes.

Orchanium is a research tool for chemists, not medical advice. It does not recommend treatments, doses or clinical use. It proposes synthetic routes backed by published evidence, and every route is a starting point for qualified chemists to verify in a properly equipped lab under normal safety practice.

## More from Meta Global AI Developer Hackathon

- [Makabeez](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/575)
- [JL](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/576)
- [saqwild](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/577)
- [IoMarkets](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/579)
- [NerdyPunkDev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/580)
- [Bellefly Kids](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/581)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/578. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt