01 / Managed
Use Tarka's standard APIs.
Approved catalog models, authentication, quotas, and usage records. A good fit for evaluations, internal tools, and workloads without special capacity or isolation requirements.
Tarka for Enterprise
AI infrastructure for workloads that need more control.
Move a production workload onto Tarka with a private model route, dedicated capacity, or a deployment designed around your existing infrastructure.
You work directly with our engineering team on model choice, performance, data handling, access, observability, and support. The result is a specific deployment plan—not a generic enterprise tier.
Deployment scope
- Model and route
- Named model versions, private endpoints, and utility access.
- Capacity and infrastructure
- Shared, dedicated, or customer-approved infrastructure according to the workload.
- Controls and support
- Scoped credentials, agreed logging and retention, and a named engineering contact.
Deployment options
Start with the least infrastructure your workload needs.
We can begin on the public platform and introduce isolation or reserved capacity when the data, traffic, or operating requirement calls for it.
02 / Private
Give your organization its own route.
Restrict a proprietary, partner, or domain-specific model to named users and credentials. Storage, payload logging, and retention are reviewed with the customer.
03 / Dedicated
Reserve capacity for the workload.
Use dedicated compute when predictable throughput, model residency, isolation, or scheduled jobs matter more than scale-to-zero economics.
04 / Customer-controlled
Run on approved infrastructure.
For organizations with existing capacity or stricter control requirements, we can review an architecture using customer-operated or approved third-party infrastructure.
Technical review
What we work through with your team
- Model quality
- Representative examples, expected answers, failure cases, and the model versions worth evaluating.
- Traffic
- Request shape, volume, concurrency, latency target, batch windows, and likely growth.
- Data
- What enters the system, where it may be processed or stored, and when it should be deleted.
- Access
- Organizations, users, credentials, model grants, operator permissions, and approval boundaries.
- Operations
- Usage records, metrics, logs, alerts, incident ownership, and changes to models or capacity.
- Commercials
- Billing currency, usage or capacity pricing, support coverage, and the commitments written into the agreement.
Typical workloads
Bring us a process your team can evaluate.
We are most useful when there is real material to test and a person who can say whether the result is good enough.
- Documents
OCR, extraction, classification, comparison, and review for forms, contracts, claims, reports, and archives.
- Knowledge
Search and question answering over approved internal material, with sources and permissions defined by the organization.
- Language and speech
Transcription, translation, and generation across Nepali, English, and multilingual content.
- Agents and batch jobs
Tool-using agents, evaluations, document pipelines, scheduled tasks, and controlled code execution.
Security and data handling
The current boundary, stated plainly.
Tarka does not use customer prompts, outputs, files, conversations, memory, or tool-call data to train models. Request and response bodies are not logged by default. If payload logging is needed for evaluation or debugging, its access and retention should be agreed before it is enabled.
Enterprise deployments can use organization-scoped credentials, customer-specific storage, explicit model grants, metadata-level usage records, approved operator access, and an agreed deletion schedule. Data-processing terms are available during review.
CURRENT CERTIFICATION STATUS
Tarka does not currently claim SOC 2, ISO 27001, HIPAA, FedRAMP, or another third-party security certification. If one of these is mandatory for your workload, tell us at the start of the conversation.
Working together
Review the workload, evaluate it, then agree the deployment.
The exact work varies, but the sequence is simple. No production commitment is implied until both teams agree the design and terms.
- 01
Review
Share the workload, representative material, traffic, data restrictions, and desired timeline.
- 02
Evaluate
Test model quality, latency, cost, and failure cases on material close to production.
- 03
Deploy
Confirm the route, capacity, access, observability, support, and commercial terms before launch.
Tell us what you want to run.
A useful note includes the workload, the model or data type, approximate volume, security or location requirements, and a target date. Rough numbers are fine.