Intrinsec AI Become a design partner

Intrinsec AI · Products

Securing the foundation of the inference stack.

  • KV cache first
  • We never see the prompt
Tessera Plato

Where they sit

Two taps on the stack you already run.

Plato taps the inference engine; Tessera taps KV storage through the KV management layer Requests enter the inference engine (vLLM or SGLang). The engine hands KV pages to the KV management layer (LMCache), which writes them to KV storage, the only copy. Plato sits at the engine and acts on what doesn't fit. Tessera sits in storage, via LMCache, and surfaces unusual pages. The prompt is never read, and storage is never copied. Requests Inference engine vLLM · SGLang KV management layer LMCache KV storage the only copy Plato At the engine Acts on what doesn't fit. Tessera In storage, via LMCache Surfaces unusual pages.
Requests become KV pages. Plato acts at the engine. Tessera surfaces in storage. Beside your stack. Never the prompt.

KV at rest

Tessera.
Surfaces unusual KV pages, wherever they live.

Watch the KV you already keep. Unusual computations surface themselves. Batteries included, and customizable.

  • LMCache
  • vLLM Production Stack
  • Dynamo
  • llm-d
lmcache server --l2-adapter '{
  "type": "plugin",
  "module_path": "tessera.taps.lmcache_l2",
  "class_name": "TesseraL2Adapter"
}'

Bottom-up discovery

Too many agent events to watch them all. So start from the computational anomalies, and work up.

KV in flight

Plato.
Acts on what doesn't fit your inference traffic.

One extra flag on the engine you already run. Off the request path.

  1. 01

    Learn

    The shape of your traffic, on any model.

  2. 02

    Surface

    What doesn't fit, tagged and classified, in real time.

  3. 03

    Act

    Through your policy. We build it with you.

  • vLLM
  • SGLang
  • Inference gateways
  • On-prem or cloud
vllm serve $MODEL --kv-transfer-config '{
  "kv_connector": "PlatonicKVConnector",
  "kv_connector_module_path": "platonic_tap.engines.vllm_connector",
  "kv_role": "kv_both",
  "kv_connector_extra_config": {
    "endpoint": "https://api.intrinsec.ai",
    "api_key": "pk_ingest_...",
    "model_id": "$MODEL"
  }
}'

The alert rides along

tool call alert policy decision alert

The alert rides along with the next event: the tool call, the policy decision. Your policy decides.

Become a design partner Watch the 2-minute demo

The 2-minute demo

See what doesn't fit.

Anomalies surfacing against an inference baseline.

The operator view

What operators see.

Traffic view: detections scored per KV page, with no prompt or response text.
Scores and labels. No prompt content.

What it covers

One baseline.

Attacks and ordinary breakage look the same from the outside. Both are deviations from how your workload normally computes.

  • Prompt injection, jailbreaks, advanced adversarial attacks
  • Sensitive-data exposure
  • Rogue and broken tool calls
  • Hidden backdoors, drift, image, audio, video

Against guardrails

Different signal. Not a bake-off.

Guardrails Plato
Reads Prompt and response text Computation
Sees the prompt? Yes Never
Per tenant Point-in-time Own baseline from the traffic
Surfaces A new rule per modality and failure One baseline

Design partners

A small number of design partners.

You bring a workload; we sit with your team until it proves itself, or doesn't.

Become a design partner