NEW A model in a box: it trains on your data, serves at sub-50ms, and the weights stay yours to keep. What's in the box →
NUMERATA
DEPLOY · NINETYFIVE

One inference engine.
Any fine-tuned model.

NinetyFive is Numerata's inference layer: sub-50ms serving for models trained on your own data, deployed inside your own trust boundary. Trade signal generation is one thing teams build on it. It isn't the only one.

PART OF THE NUMERATA STACK  ·  ninetyfive.gg ↗

CORE CAPABILITIES

Sub-50ms latency, for whatever you're serving.

Request and response latency Your app sends a request to NinetyFive and receives a response in under 50 milliseconds. YOUR APP NINETYFIVE REQUEST RESPONSE IN UNDER 50MS
Same round trip your app already makes. Just fast enough to not notice.

CUSTOM MODEL SERVING

Serve models fine-tuned on your own data, not a generic public endpoint, so every response matches your domain, not someone else's.

SUB-50MS LATENCY

Built for real-time use cases: trading signals, risk scoring, chat and agents, anything where response time is part of the product experience.

PRIVATE BY DEFAULT

Runs inside your own infrastructure, private cloud or fully air-gapped. Fully offline local inference is coming soon.

BUILT FOR

The same low-latency layer, wherever you point it.

Real-time trade signal generation, scored on your own models, is proof this layer holds up under one of the tightest latency budgets we could find. Research assistants, compliance review agents, and client-facing chat run on the exact same serving path, just pointed at a different model.

One inference engine, an open-ended set of use cases A single NinetyFive box on the left connects through a shared vertical bus to trade signal generation, research assistants, and compliance review agents, plus a fourth dashed row labeled whatever you build, indicating the list isn't exhaustive. NINETYFIVE INFERENCE TRADE SIGNAL GENERATION RESEARCH ASSISTANTS COMPLIANCE REVIEW AGENTS + WHATEVER YOU BUILD
Same engine, same latency budget. These are examples, not the full list.

YOUR DATA STAYS YOURS

Everything sent to NinetyFive is time-boxed, and never used to train any model except your own, critical when the data is trade or client data.

Fourteen-day retention window Data comes in, sits inside a fourteen-day retention window, and is then automatically deleted rather than kept or used to train any other model. DATA IN 14-DAY RETENTION WINDOW DELETED
Never routed to any model except your own, and never kept past the window.

For code-based deployments specifically, binary files, anything covered by .gitignore, and files processed with git-crypt are excluded automatically. Teams that need more can run secure compute and secure training entirely within their own infrastructure, the same private-by-design posture behind every layer of Numerata's stack.

FURTHER READING

More on the serving layer.

GET STARTED

Deploy your own fine-tuned models today.