built a thing that finds exact answers in your documents without the search service ever seeing your text.
How it works: you paste your documents and your question into the browser. They get encoded locally into a numeric field envelope (client-side encoding operates at $O(N)$). That envelope gets sent to my remote oracle, Gravitas, which performs the core search calculation in $O(1)$ time. You decode the answer locally. Exact answer, or nothing. No hallucinations. No ranked list. No probabilities.
The system does not rely on phoning home. The service never receives raw text, queries, or metadata, acting purely as a zero-trust validator for signed int16 arrays. Every transaction produces a deterministic proof receipt.
I ran it against MS MARCO v2.1 (138M passages, 285K queries) and MLEB (all legal tracks). Perfect fidelity on the commercial encoder. The protocol is open-source (OpenEncoder, Apache 2.0) so anyone can audit the wire format.
Live site: https://mushku.com. While it serves as a live demo, there is a pay-per-search function built in so you can get instant, custom access to the oracle for your own data right now.I am a normal person who made a cool thing. I don't know how to do marketing. If you want to talk about it, email me at sram@mushku.com. What should I build next?
built a thing that finds exact answers in your documents without the search service ever seeing your text. How it works: you paste your documents and your question into the browser. They get encoded locally into a numeric field envelope (client-side encoding operates at $O(N)$). That envelope gets sent to my remote oracle, Gravitas, which performs the core search calculation in $O(1)$ time. You decode the answer locally. Exact answer, or nothing. No hallucinations. No ranked list. No probabilities. The system does not rely on phoning home. The service never receives raw text, queries, or metadata, acting purely as a zero-trust validator for signed int16 arrays. Every transaction produces a deterministic proof receipt. I ran it against MS MARCO v2.1 (138M passages, 285K queries) and MLEB (all legal tracks). Perfect fidelity on the commercial encoder. The protocol is open-source (OpenEncoder, Apache 2.0) so anyone can audit the wire format. Live site: https://mushku.com. While it serves as a live demo, there is a pay-per-search function built in so you can get instant, custom access to the oracle for your own data right now.I am a normal person who made a cool thing. I don't know how to do marketing. If you want to talk about it, email me at sram@mushku.com. What should I build next?