Purpose-built LLMs, running where, when and how YOU determine.

Signal Stable is a managed AI platform from Kohala Tech. Select, configure, deploy, and support domain-specific AI systems that live inside your infrastructure boundary — not ours, not a frontier vendor's. Right-sized capability. Full and verifiable control. Pay-as-you-go economics, scale-to-zero when idle.

Sign up for access → Explore the stack
The Platform

Not a raw model. A production platform.

Signal Stable is the managed layer between you and open-weight AI. Kohala selects the right model for your work, wraps it in a guardrailed runtime, deploys it inside the AWS environment of your choice — ours or yours — and operates it as a service. You get a production-ready system — not a weights file to figure out yourself.

Flexible infrastructure
Hosted by us, or in your cloud
Run it in a Kohala-managed dedicated environment we operate end-to-end, or deploy directly into your AWS account — we handle the cross-account wiring either way. Single-tenant. No shared compute with other customers.
Production runtime
Guardrailed from day one
Anti-hallucination scoring, RAG-first retrieval, structured JSON output, source attribution, uncertainty flags. The things every serious LLM deployment eventually builds — you get them on day one.
Multi-tenant
One deploy, many workloads
Runtime domain switching lets a single deployment serve separate teams, projects, or business units with isolated context, prompts, and audit trails — without provisioning new infrastructure for each.
Elastic economics
Pay-as-you-go, scale to zero
Billed by the compute-hour you actually consume. No monthly minimums, no reserved-capacity bets. When your tenant isn't running a query, your spend is zero.
The Nova Stack

Four tiers. One product layer. Your terms.

Every Nova tier runs single-tenant — in your AWS account or in a Kohala-managed environment — wrapped in the Kohala product layer: anti-hallucination, RAG-first, audited, multi-tenant. Pick the tier your workload actually needs; scale to zero when it's idle.

Privacy by design · Every tier
Strong tenant isolation. Unparalleled privacy. Your prompts, your data, your outputs never leave your boundary.
Your dedicated VPC Your keys Your policies
Flexibility scale-to-zero · pay-as-you-go
Contact sales to get a quote today →
Nova-S
Fast + Cheap
Built for
  • Document classification at scale
  • Structured extraction from PDFs & forms
  • Simple Q&A over known corpora
  • High-volume, low-latency production workloads
Capability
Fast, efficient, cost-predictable. The production tier for well-defined, high-volume tasks — classification, extraction, and structured output.
Nova-M
Production workhorse
Built for
  • Scholarly research & evidence synthesis
  • Regulatory analysis & compliance reporting
  • Multi-turn operational advisory
  • Runs the life sciences domain package
Capability
The balanced workhorse. Strong enough for evidence-grounded synthesis, regulatory analysis, and multi-turn advisory — at a price and speed built for real production scale.
Nova-L
Near-frontier reasoning
Built for
  • Complex legal & financial analysis
  • Multi-document synthesis at scale
  • Strategic advisory with nuanced reasoning
  • Extended context — swallow entire contracts
Capability
Deeper reasoning, extended context. Built for the nuanced work where the quality of the answer matters more than the speed of the response.
Nova-XL
Frontier-class, in your VPC
Built for
  • Frontier-grade multi-step reasoning
  • Full-corpus analysis in one pass
  • Multilingual enterprise workloads
  • When private deployment AND frontier-class depth both matter
Capability
The top of the Nova lineup. The most capable tier — for the heaviest multi-step reasoning, the longest documents, and the most demanding multilingual workloads.
Every Nova tier ships with the Kohala product layer
Frontier Augmentation

Thoughtfully augmented with the most powerful models on the planet.

Your Nova tier handles the vast majority of work. When a request genuinely needs frontier depth, Anthropic's Claude steps in — through your policies, your guardrails, your audit trail. You control where your information flows.

◆ Available on demand
Claude, when you need it.
Anthropic Sonnet & Opus, integrated into the Signal Stable platform
How it works
Routes through the Inference Broker. Nova-first by default. Claude engages only when the broker classifies the request as exceeding Nova's capability — or when you tag it explicitly. Every escalation logged. PHI stripped before anything leaves your boundary.

Frontier quality, on your terms. The most capable models on the planet, available through the same Signal Stable control plane as your in-VPC Nova workloads — governed by your policies, logged in your audit trail, scoped to the requests you approve. You surface the capability. You keep the control.

What you get
  • ⇌
    Broker-routed, not direct
    Tag-based + complexity-classified. Nova-first by default; Claude only when the broker says it's warranted.
  • 🔒
    PHI / PII redaction pre-escalation
    Sensitive tokens stripped before any request leaves your infrastructure boundary. Logged, reversible, auditable.
  • 📊
    One platform, one audit trail
    Every request — Nova and frontier — lives in the same log, under the same policy surface. No switching consoles, no stitching reports.
  • ◈
    Frontier reasoning on demand
    Tag a request, hit a complexity threshold, or let the broker decide. The deepest model you need, only when you need it.
  • ✕
    Opt-out at any layer
    Tenant, domain, or request. Turn frontier escalation off and every call stays inside your boundary. You retain the choice.
Domain Package · Life Sciences

A life sciences LLM, trained for researchers.

Our life sciences domain package is Nova-M trained and tuned for evidence-based research: PubMed-indexed retrieval, GRADE/CEBM evidence grading, clinical calculators, HIPAA-grade audit. Built for life sciences researchers and clinical teams; generalizable to any academic medical program, biotech, or research group that wants the rigor of a scholarly assistant without shipping PHI to a frontier vendor. Customers deploy it under whatever name fits their institution.

◆ Domain package · Life sciences
Life Sciences Package
Our reference build — customers deploy under their own name
Availability
Only available on Nova-M or base model. Domain quality requires the 14B minimum.

The Life Sciences package is Nova-M running a life sciences system prompt overlay, three curated RAG collections, and a toolkit scoped to research workflow. Every response is evidence-graded, source-cited, and PHI-aware. Claude escalation is available for the complex questions that exceed Nova-M's capability — with PHI stripped first.

What's in the package
  • 📚
    Three RAG collections
    PubMed / MEDLINE · Clinical practice guidelines · Discipline-specific protocols
  • ⚖︎
    Evidence grading
    GRADE + CEBM quality assessment on every claim. Never asserts beyond the evidence.
  • 🔬
    Clinical toolkit
    PubMed search (MeSH-expanded), guideline lookup, clinical calculators (Wells, CHADS-VASc, APACHE-II, qSOFA)
  • 🛡︎
    HIPAA-grade controls
    Automatic PHI detection · de-identification before cloud escalation · IRB flagging on human-subjects research · enhanced audit trail
  • 📝
    Four output templates
    Literature review · evidence summary · clinical answer · protocol draft — each with citations and evidence grades
Trust & Controls

Private by architecture, not policy.

Your security team's first question will be some variant of "can this pass review?" Here's what we put in front of them.

Data privacy
Your prompts and outputs never leave your dedicated VPC.
  • All model inference runs inside your dedicated VPC — your AWS account or a Kohala-managed environment
  • No training-on-your-data — ever
  • KMS-encrypted at rest, TLS in transit
  • Kohala staff cannot read into your data plane
Audit posture
Every request, traceable end-to-end.
  • CloudTrail integration for all control-plane actions
  • Per-request audit log: user, prompt, retrieval sources, output
  • Tamper-evident retention in your S3 bucket
  • Exportable to your SIEM (Splunk, Datadog, etc.)
Platform telemetry (what we do see)
Uptime and health, not content.
  • Instance health + ALB metrics (so we can keep it running)
  • Latency and throughput aggregates
  • Error rates and patch-compliance posture
  • Zero visibility into prompt content or user data
Governance
Model weights and updates on your terms.
  • Open-weight base models with verifiable provenance
  • No silent model updates — versioned, opt-in upgrades
  • Policy engine: your allowed topics, blocked prompts, data classifications
  • Air-gap deployment available for regulated workloads
Need a specific compliance framework (HIPAA, SOC 2, FedRAMP alignment)? Talk to us →