Persistent, permission-aware context
Manage context across tasks, with design goals for provenance, correction, access, and retention. This means stored context, not automatic model retraining.
Explore the Membrane direction →Public roadmap
Execution roadmap.
Built today, validating now, and planned next. Srasta separates current capability from planned product direction, and every public claim is tied to a deployment/runtime profile we can actually prove.
Roadmap discipline
We do not pitch roadmap items as shipped capability.
Deployment support follows the published catalog. Model access, knowledge retrieval, access controls, and operations form the current platform. Additional deployment profiles and planned products need their own validation before becoming available.
Planned · Not available today
Memory and agents for company workflows.
These are separate product directions. Scope, availability, and pricing are not committed; current evaluations should stand on available capabilities.
A foundation for your own agents
Help companies build agents using models, tools, and memory. We are gathering requirements for permissions, human approval, execution boundaries, observability, and recovery. No SDK, integration list, or release date is committed.
Discuss an agent workflow →Milestones
Where the platform evolves.
The current deployment rows below follow the support catalog. Broader deployment profiles and product directions require validation before they become supported capabilities.
Trial/license funnel, private inference, governed RAG, admin, install control, audit foundations, license ops, catalog studio, and security review packet.
BuiltCertified now
BuiltCertified now
BuiltCertified now
BuiltHugging Face and Unsloth source intake, reviewed candidates, runtime profiles, evidence capture, and publish workflow are being hardened for operator use.
ValidatingTighten clean-room reset, provenance, service discovery, audit delivery, user onboarding, and proof receipts so validation is repeatable without stale local state.
NextNVIDIA multi-host, production operations, backup/restore, rollback, support bundles, and private registry posture.
NextKubernetes, HA, AWS, Azure, GCP, Vast, and other runtime adapters become supported only after their own certification evidence exists.
LaterPlanned Membrane and agent harness work will be shaped by customer requirements. Scope, dates, and pricing are not committed.
LaterBuild detail
What each phase includes.
Private AI foundation
The GTM baseline is built: private inference, governed RAG, admin operations, license lifecycle, role-aware model access, audit foundations, and customer-controlled deployment profiles.
- Private inference path for open-weight models on customer-controlled hardware
- Gateway foundations for role-aware model access and auditable AI requests
- Admin foundations for users, roles, licenses, model access, and runtime visibility
- Audit foundations for model access and administrative activity
- Catalog authority for which hardware, runtime adapters, inference engines, and model bindings are installable
- Srasta-Agent heartbeat, runtime health, controlled actions, and proof receipts
Certified deployment profiles
Current certified profiles: Blackwell Gb10 / vLLM single node, Apple Silicon / MLX single node, Blackwell Gb10 / vLLM single node. Support follows the published catalog for the selected hardware, runtime, and model.
- Check the supported-deployment matrix before choosing hardware
- Confirm the selected model and runtime combination for your workflow
- Additional configurations require their own validation evidence
Catalog certification factory
Catalog Studio is becoming the operator workflow for adding and certifying more hardware, inference engines, embedding models, runtime profiles, and LLM models without changing customer-side code.
- Hugging Face and Unsloth source intake for model candidate discovery
- Operator review screens for model metadata, runtime fit, quantization, parser support, and evidence
- Certification targets for Apple MLX and NVIDIA vLLM profiles
- Publish workflow that updates the catalog authority consumed by license and installer flows
- Public supported-deployment matrix sourced from the same catalog data
Repeatability and day-2 hardening
The next hardening gate is repeatable proof from clean state: no stale local data, no hidden fallbacks, no duplicate proof paths, and no unsupported service discovery assumptions.
- Clean-room install reset and provenance checks
- Service discovery as the only installed-platform service authority
- Audit authority delivery and replay hardening
- User onboarding, activation, invite, and recovery regression coverage
- Support bundles and operator-visible proof receipts
Multi-host, Kubernetes, and HA operations
Multi-host GPU, Kubernetes, HA, cloud-provider adapters, backup/restore, rolling upgrade, registry trust, support collection, and SLA posture require their own validation evidence before becoming supported profiles.
- NVIDIA multi-host deployment profile
- Kubernetes operator and upgrade posture for enterprise deployment profiles
- HA-oriented topology guidance for gateway, inference, admin, governance, memory, and audit layers
- No-planned-downtime upgrade patterns where architecture and customer infrastructure allow it
- Disaster recovery, restore drills, rollback drills, and proof artifacts
- Formal SLA commitments only after measured validation with paying customers
Governed integrations, Membrane, and specialized enterprise requirements
These are not day-one promises. We will scope and price them when a paying customer has the requirement and the environment to validate it.
- Governed Tool Pack registry and approval lifecycle
- Agentic harness support for customer workflows that need planning, tool use, and operator guardrails
- Membrane runtime for governed memory and reusable organizational intelligence
- Customer-specific certified images and private registry
- Windows node support
- Vast.ai or similar GPU-marketplace customer installs
- Formal five-nines SLA backed by measured production evidence
- Advanced workspace, role, and sector-specific compliance automation
- AMD, Intel, and specialized edge AI machines after demand and validation
Design partner fit
Have a workflow that should shape the roadmap?
The best roadmap input is a real environment, a real governance boundary, and a measurable workflow. We capture requests in the same CRM-backed funnel as pilot and product-update interest.