# Team Orionmatter

- **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:** [Aditya Sahare](https://cerebralvalley.ai/u/orionmatter)
- **GitHub:** https://github.com/experiments-of-heisenberg/orionmatter-path-engine
- **Demo video:** https://youtu.be/J1vx0-ercSA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3898

Orionmatter Path Engine is a career route planner for engineers that shows its work.

The problem: ask a chatbot how long it takes to reach Staff Engineer and you get one confident answer with no range, no sample size and no evidence. Engineers make real decisions (switch companies, retrain, relocate, share their story) on advice like that.

What it does: you describe your career in your own words, or upload a cropped screenshot. Orionmatter shows three routes to your goal (climb where you are, jump to another company, pivot to another path). Each route is a range, never a promise, with the number of comparable engineers behind it and its evidence.

Five agents run on Muse Spark, with one rule: the model writes the words, code owns every number and every verdict.
- Intake reads your story, or a screenshot using vision. Each field shows its source. For typed text, code checks that the quoted words are really in your story. Nothing is used until you confirm it.
- Planner is a tool-calling agent. The model chooses which evidence to look up, code runs each lookup, and every result gets an evidence ID that the plan must cite.
- Skeptic is a code verifier plus a model critic. It checks every figure in the plan against the evidence. In our recorded live run it caught a claim of n=55 where the evidence said n=53 and sent the plan back with the exact words. Only an approved plan puts numbered steps on your map. A plan that stays rejected is shown as unresolved, never hidden.
- What-if turns a sentence ("what if I want to get there within two years?") into limit changes. Code recomputes every route, and the model's explanation is rejected unless every figure matches what code computed.
- Privacy Guard shows how many comparable engineers still match as you reveal role, level, goal and region (197, 93, 68, then 11). It accepts a model-proposed finding only if the quote is really in your text, and the share card keeps you in a crowd of ten or more.

Tested like a product: nearly 2,000 automated tests (fixes confirmed by reverting them), secret scanning in CI, and live evaluations of the real agents: intake 18/18, interpreter 28/28, prompt-injection checks 14/14. These are small suites, so they mean "no failures in these cases", not "never fails".

Honest limits: the cohort of 484 engineers is synthetic, so routes show how the method behaves, not how real careers go. Pay is an indicative market range, not cohort data. It is a planning tool for one person, never for screening candidates.

Live app: https://orionmatter.vercel.app
Demo (no key needed): https://orionmatter-demo.vercel.app
Code: https://github.com/experiments-of-heisenberg/orionmatter-path-engine
Video: https://youtu.be/J1vx0-ercSA

## More from Meta Global AI Developer Hackathon

- [Builders101](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3895)
- [Solverine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3896)
- [Zeynep Sude Kilecioğlu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3897)
- [TinyTMS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3899)
- [ZendragonVR](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3900)
- [Strata](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3901)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3898. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt