- Hardware and software delivered as one unit, integrated before it reaches your rack
- Tuned for sustained load rather than benchmark peaks
- Surogate ships with it, ready for agent and training workloads
- Up to 8× NVIDIA RTX 5090 or RTX PRO 6000 Blackwell GPUs
New modelSurogate 3.7 Flash
BUILD.DEPLOY.OPERATERELIABLE AI
The AI control plane for your infrastructure, your data, your models and your agents.
Banks, regulators, telecoms and the state run on it.
31clients and partners
All of themPilots are easy. Production is not.
Enterprises have plenty of agent prototypes. What they lack are agents that survive contact with production: real data, real volume, and real consequences when a decision goes wrong.
What is missing is the operational ground under the model: identity, policy, traces, evaluation, cost control, and a way to be better than the day you shipped.
Where to start
The appliance underneath, the platform on top and the automation layer around them. Start with the one your teams need first.
- Document workflows, process automation and the long tail of work between systems
- A foundation your teams extend without writing code
- Mission-critical KPIs visible the moment they move
- End-to-end data governance, from APIs through to storage
- Surogate 3.7 Flash and our other models, trained in our lab for agent work
- Surogate Agents: build, govern and run agents for whole teams, without code
- Keel: run agents inside processes, under rules and approval limits you own
- AI Gateway: every model behind one API, with keys, budgets and failover
One control plane for the whole agent lifecycle
Build, run, observe, evaluate, govern and improve agents in one place, on infrastructure you control. Agent definitions, skills, tools, models and datasets are versioned in one registry, so what runs in production is exactly what you approved.
Explore Surogate- Build
Compose agents from skills, tools, MCP integrations and sub-agents. Your domain experts write the skill in plain language. The platform turns it into capability an agent can execute and an evaluation can score.
- Deploy
Ship agents as versioned artifacts on Kubernetes, on your servers, in your cloud or air-gapped. Every rollout repeats the last one exactly.
- Operate
Every run leaves a full execution trace. Replay any session step by step. Guardrails, budget caps and audit logs are enforced while the agent runs.
- Improve
Production traces become training data. Training data becomes specialized models. Approved models return to production, and the agents get measurably better the longer they run.
Proven where the consequences are real
We deploy inside your data boundary, your compliance regime and your operational reality. These are the sectors we already work in.
Finance
01Fraud moves faster than a review queue, and the regulator’s questions do not wait.
- Real-time fraud detection
- Automated risk and compliance reporting
- Client insight distilled from your own data
- Faster reconciliation and audit workflows
Telecom
02Networks the size of a country, and subscribers who notice every second of degradation.
- Predictive network optimization
- Support agents that resolve rather than deflect
- Churn prediction and retention
- Automated billing and revenue assurance
Healthcare
03Clinical data is deep, sensitive, and never allowed to leave.
- Clinical decision support for diagnostics
- Accelerated medical imaging analysis
- Scheduling and intake assistants
- On-premise deployment for HIPAA and GDPR
Manufacturing
04Downtime, defects and supply volatility compound.
- Predictive maintenance of equipment
- Automated defect detection in production
- Demand forecasting and supply chain planning
- Worker safety monitoring
Retail
05The shelf, the app and the call centre are expected to know the same thing at the same moment.
- Personalized recommendations
- Support that knows the order history
- Demand forecasting
- Loss prevention
- Continuity across every channel
Public Sector
06Constrained budgets, rising expectations, and a sovereignty requirement that is not negotiable.
- Citizen service delivery
- Public safety
- Social services casework
- Infrastructure management
- Policy analysis and research
What specialization is worth
A compact model, trained on one agent’s task from that agent’s own production traces, measured against a general frontier API on the same evaluation set.
Your specialized model95.8%
Frontier API94.1%
Your specialized model380 ms
Frontier API2,400 ms
Your specialized model$0.40
Frontier API$9.20
Our own measurements. Yours will differ with the task, so every deployment starts by building an evaluation set of your own.
< 30 days
From proof of concept to production
99.95%
Availability SLA on DenseMAX appliances
< 12 months
To return on the investment
Where intelligence is built
From infrastructure to LLM architectures, custom serving kernels and our own training framework, we advance AI development through continuous research and innovation.
Research
Agent and skill optimization through RL and self-evolving harnesses
Custom serving engines for Blackwell, Ada, Hopper and ROCm
High-speed training algorithms for SFT and GRPO on consumer cards
Open source
Trained models
- Surogate 3.7 Flash, our language model for agent work, with 1M tokens of context
- Surogate Rune, a decision model for text and images
- Surogate Speech, jackrabbit for speech to text, Amami for text to speech
Expert models fine-tuned on a client’s own production traces
- Check out our HuggingFace page
Enforced, not documented
Every agent runs inside a boundary you drew. Identity, budget, redaction and audit are enforced while the agent runs.
Zero-trust defaults
Role-based access control, project isolation, scoped credentials in a vault, network isolation.
Agent-scoped authority
Every agent is a registered, versioned object with its own credentials and its own budget cap. Nothing is inherited from whoever launched it.
Data residency
Your VPC or your own metal. Your data, your traces and the weights you train stay inside your perimeter.
Policy guardrails
Redaction, PII handling, content filters, throttling and automated red-teaming, applied to the agent as it runs.
Operational hardening
Encrypted at rest and in transit, secure boot, TPM attestation, sandboxed execution, resource isolation.
Lineage and rollback
Versioned artifacts, dataset cards and reproducible fine-tuning. Every result traces back to the run that produced it.
ISO 27001SOC 2 Type II (in progress)GDPR / Schrems IIHIPAA-ready architectures
Sovereign by default
The same platform, wherever your data is allowed to be.
01
DenseMAX appliance
On your floor, on your power. Hardware and platform delivered as one and tuned together.
02
Your own cloud
AWS, GCP, Azure or Oracle Cloud, inside your VPC. Kubernetes-native, multi-node, autoscaling.
03
Hybrid
Train on the appliance and burst to the cloud, or the reverse. One control plane across both.
04
Air-gapped
Fully isolated deployment for environments where nothing leaves, including the models you trained.
Open by construction: Kubernetes for the substrate, the Model Context Protocol for tools, OpenAI-compatible endpoints for serving, HuggingFace and ModelScope for import and export, and an open-source training core. Your agents, your datasets and your trained models are portable out of here on the day you want them to be.
Bring us a workflow that matters
One real use case and your infrastructure constraints are enough to start. We will show you the path to production, how to govern it and what it costs to run.