RESEARCH

Engineering notes

These are SentryNeural's engineering notes: short write-ups on how we think about building AI and cloud systems that are safe, reliable and easy to run.

Design notes

Each note describes an approach we use or are developing. They are design notes and targets, not measured results, and none has been peer-reviewed.

SN-RFC-01 ● Design note

Deterministic JSON Grammar Enforcement at Inference Ingress

Describes a compiler that turns JSON schema contracts into filters applied while a model generates its answer, so outputs are designed to follow the agreed structure. The aim is a small added delay per request, which has not yet been measured.

Grammar enforcementvLLM inferencePredictable outputs
SN-RFC-02 ● Design note

Zero-Data-Retention Memory Sandboxing in Ray and vLLM Clusters

Architectural patterns for running AI workloads only in volatile memory, designed to avoid sensitive data being written to disk and to erase session context as soon as a request completes.

Memory onlyRay clusterSecure erase
SN-RFC-03 ● Design note

Adversarial Prompt Hijacking Interception via Token-Level Bounding

Evaluates streaming filters placed ahead of the main model to catch attempts to hijack it, such as role-inversion, indirect jailbreaks and attempts to extract sensitive settings.

Ingress boundaryJailbreak interceptionLow added delay

Any figures we publish in future will come with the method used to produce them. Until then, treat the notes as design intent.

Want to talk about your own project?

Tell us about it in a free discovery call. You will talk to the founder, not a sales team.

Book a free discovery call

Or email contact@sentryneural.com