Workloads
Production-ready encrypted apps
AI inference, vector search, and other reference workloads you can deploy as-is or extend.
AI inference and database queries on encrypted data, at cloud
scale, with zero plaintext exposure.
THE PRIVACY PROBLEM
Every existing approach asks you to give something up: cloud scale, data utility, or trust in the hardware. We built Lattica so you don't have to choose.
| Approach | Cloud scale | Data utility | Zero trust |
|---|---|---|---|
| On-Prem | |||
| Anonymization | |||
| Confidential computing | |||
| Lattica |
THE LATTICA APPROACH
FHE lets a server compute on data it can never see. The input, the output, and everything in between stay encrypted. The math has been understood for years; the hard part is making it fast enough to actually use.
Lattica builds the full stack, cryptography, compiler, and GPU runtime, as a single system designed end-to-end so encrypted workloads run at the speed real applications need.
AI inference, vector search, and other reference workloads you can deploy as-is or extend.
Queries are encrypted in the user's environment, sent to the Lattica API, and executed without ever being decrypted.
The cryptography, compiler, and GPU kernels that make encrypted compute fast enough for production.
HOW IT WORKS
A single platform for deploying and querying encrypted workloads. Service providers ship models and databases as-is; end users integrate with a lightweight client and a familiar API.
AI models or databases are uploaded to the platform once, ready to serve encrypted traffic.
Queries are encrypted locally before anything leaves the user's device.
GPU-accelerated FHE runs the workload directly on encrypted data. No plaintext, ever.
The encrypted result returns to the user, who alone holds the decryption key.
DEVELOPER LIBRARY
Prototype to production, without FHE complexity. Compose encrypted pipelines using familiar tensor and model patterns, then deploy them to Lattica Cloud with a single call.
Compose encrypted pipelines from tensor ops, model layers, and client-side reshapes - no cryptography in your code path.
The compiler picks parameters, packing strategies, and kernel schedules so your workload runs correctly and fast by default.
Ship to Lattica Cloud with a single deploy(), pick a hardware profile that matches your latency and cost targets.
Built-in workload access control, versioning, and observability for the workloads you expose to end users.
from lattica.deployer.hom_ops import SequentialHomOp, HomLinear, HomSquare
from lattica.deployer.client_ops import ClientReshape
from lattica.deployer.admin import admin_api
hom_mnist = SequentialHomOp(
# preprocess: flatten 28x28 input image to a 784-d vector
ClientReshape((BATCH_SIZE, 28 * 28,)),
# first linear layer
HomLinear(l1_weight.shape),
# square activation
HomSquare(),
# second linear layer
HomLinear(l2_weight.shape),
# postprocess: reshape output to 10-class vector
ClientReshape((BATCH_SIZE, 10,)),
)
admin_api.deploy(hom_mnist, n=2**14, q_bits=124, device='gpu-L4-aws')WHERE LATTICA FITS
The same execution layer powers very different workloads: from regulated inference inside banks and hospitals, to hosted models serving enterprise customers, to lookup services where even the question is sensitive.
Explore reference workloads01 · Confidential inference at scale
Run your own models on data too sensitive for the cloud.
Financial services
Risk · Credit · Fraud scoring
Healthcare
Clinical risk · Claims · Fraud
Pharma
Trial screening · Safety signals
02 · Hosted models, private customer data
Sell inference to enterprises without ever touching their data.
Fraud & risk
Transaction · ATO · Chargeback
Medical AI
Imaging · Triage · Lab analysis
Identity & verification
Biometrics · Liveness · Docs
03 · Private lookups & registries
Answer queries without learning what was asked.
Caller ID & comms
Spam · Caller identity
Threat intelligence
Malware · Phishing · Domains
Sanctions & compliance
AML · PEP · Sanctions