One-line pitch: We split sensitive text across two models: a local encoder and a remote specialist that works on vectors, not your sentence. Sealed obfuscates vectors on the wire; HELIX proves encrypted specialty routing where the math fits. Full encrypted generation is future work.
Research demo — not a clinical product. Based on Gorbett & Jana, 2025.
| Doc | Use when |
|---|---|
| BUILD.md | Install, download models, train alignment, run server + UI |
| hackathon_docs/overview.md | Understand architecture, privacy ladder, limits |
| hackathon_docs/pitch.md | June 6 judges — 3 min pitch, 7 slides, submission |
| hackathon_docs/demo-roadmap.md | Longer live demo (lab / backup) |
| hackathon_docs/spickzettel.md | Cheat sheet — what runs where, alignment, LM head, Sealed, HELIX/CKKS |
| hackathon_docs/use_cases.md | Other industries (pilot sketches) |
Quick run (after BUILD.md)
./demo_split_all.sh
# Clinic → http://localhost:4200
# Hospital → http://localhost:4201
# Passphrase both sides: hackathon2026| Step | Real gain |
|---|---|
| Prompt stays on clinic | Data minimization — not encryption alone |
| Alignment | Maps spaces; hospital still gets vectors |
| Sealed (rotation) | Wire obfuscation — server decrypts for generation |
| HELIX (CKKS) | Crypto for 5-class routing only (~3–4 s CPU) |
Details: overview.md.
Gorbett, T., & Jana, S. (2025). Characterizing Linear Alignment Across Language Models. arXiv:2603.18908.