Six AI workloads.
Zero plaintext.

Retrieval, agent memory, biometric search, multi-tenant isolation, cross-org similarity, and private embeddings — all running on the same encrypted index.

USE CASE / 01

RAG without plaintext in flight.

Retrieval runs entirely on ciphertext. The LLM only sees what your client decrypts; we never see anything. Provider-agnostic — pair with Anthropic, OpenAI, or your own model.

No plaintext in flight
Provider-agnostic
Audit-ready
Optional attested TEE inference

QUERY PIPELINE · CIPHERTEXT IN-FLIGHT

STAGE 01

"site policy?"

Query

STAGE 02

7AD0·369C

Encrypt

STAGE 03

E147·AD03

Top-k

STAGE 04

"with: doc-7.."

LLM prompt

CLIENTEDGECYBORGDBLLM

USE CASE / 02

Agent memory with a cryptographic audit trail.

Every retrieval an agent makes is signed and appended to a tamper-evident ledger. Different users decrypt different views of the same memory — without splitting the store.

Append-only signed log
Per-identity decryption keys
Replay any agent session
Compliance-grade evidence

SIGNED LEDGER · APPEND-ONLY

#0001 agent.researcher · retrieve · policy/2026-Q1 ✓ sealed

#0002 agent.summarizer · retrieve · minutes/board ✓ sealed

#0003 user.alice · decrypt · policy/2026-Q1 ✓ sealed

#0004 agent.draft · retrieve · contracts/acme ✓ sealed

#0005 user.bob · decrypt · contracts/acme ✓ sealed

#0006 agent.tools · retrieve · invoices/Q4 ✓ sealed

#0007 user.alice · rotate-key · policy/2026-Q1 ✓ signing…

USE CASE / 03 NEW

Biometric search on encrypted templates.

1:N identification against galleries of 100M+ subjects, sub-1Ca0ms — and no plaintext template is ever reconstituted. Compromise? Rotate the key. No re-enrollment needed.

1:N on ciphertext
Revocable enrollments
BIPA / GDPR Art. 9 ready
Liveness-aware

1:N IDENTIFICATION · GALLERY = 10M

match

✓

· · ·

USE CASE / 04

Multi-tenant search without shared keys.

Each tenant gets its own KEK. Cross-tenant queries are not possible — not because a `WHERE` clause says so, but because the keys to read them do not exist in the same place.

One KEK per tenant
No shared decrypt path
Per-tenant key rotation
Operator can't escalate

CRYPTOGRAPHIC ISOLATION · 1 KEK PER TENANT

ACME

8 vectors · sealed

GLOBEX

12 vectors · sealed

INITECH

6 vectors · sealed

UMBRELLA

10 vectors · sealed

— no shared keys · no cross-tenant decrypt path —

USE CASE / 05

Cross-org similarity without shared records.

Private set intersection on encrypted vectors. Both sides learn only the matches; no record, embedding, or identifier crosses the boundary. Useful for fraud rings, sanctions, joint diligence.

Encrypted PSI
No raw records exchanged
Threshold-tunable matches
Query audit trail

PRIVATE SET INTERSECTION

USE CASE / 06

Embeddings the server never sees.

Embedding can be generated client-side or in a TEE; only ciphertext leaves your device. The index never holds a plaintext vector.

Client- or TEE-side embedding
Plaintext never egresses
Drop-in for OpenAI / Cohere clients
Same recall as plaintext

ENCRYPT-THEN-EMBED · NO PLAINTEXT EGRESS

Six AI workloads.Zero plaintext.