Platform runtime
Clustering. Peer cluster with gossip membership.
Sharding. Deterministic entity partitioning.
Entity rebalancing. Redistributes on membership change.
Durable execution. Workflows resume from last step.
State replication. Entity state across zones.
Multi-region. Active-active state across regions.
Split-brain resolution. Configurable partition strategies.
HA / DR. No single point of failure.
Rolling updates. Zero-downtime version rollout.
Auto-scaling. Cluster size follows load.
Self-healing. Failed members replaced automatically.
Backpressure. End-to-end flow control.
Work-stealing. Idle members pick up queued work.
Gossip protocol. Peer-to-peer membership.
Failure detection. Phi accrual, tunable sensitivity.
Cross-region failover. Reroutes on region degrade.
Session stickiness. Requests routed to entity shard.
In-memory grid. State in RAM for microsecond reads.
Akka Actors
Actor model. Unit of state, behavior, and one mailbox.
Serialized execution. One message at a time, no locks.
4KB actors. Memory footprint of an idle actor.
200M actors per core. Density on a single CPU core.
1M-node clusters. Cluster size the runtime is built for.
Actor hierarchy. Parents supervise the children they create.
Supervision strategies. Resume, restart, stop, escalate.
Backoff supervision. Restarts spaced by backoff.
Death watch. Termination arrives as a message.
Typed behaviors. Protocols checked at compile time.
Tell and ask. Fire-and-forget, or request with reply.
Mailbox types. Unbounded, bounded, priority, control-aware.
Stashing. Hold messages until the actor changes behavior.
Dispatchers. Thread-pool, pinned, affinity pools.
Actor timers. Single-shot and periodic messages.
Receive timeout. Fires when an actor idles past a threshold.
Routers. Pool and group, round-robin or hashed.
Location transparency. One reference reaches any node.
Remoting. Actor messaging over encrypted transport.
Cluster singleton. One instance across the cluster.
Distributed pub-sub. Topic delivery across the cluster.
Circuit breaker. Trips a failing call path, half-opens to retry.
Coordinated shutdown. Ordered phases on leave or exit.
Stream integration. Actors as stream sources and sinks.
Development model
Agent component. LLM unit with session memory.
Workflow component. Durable multi-step orchestration.
Event-sourced entity. State from event log.
Value entity. State as latest value.
View component. Materialized projection over events.
Endpoint component. HTTP, gRPC, or MCP surface.
Consumer component. Subscribes to events or topics.
Timed action. Scheduled durable work.
Component composition. Typed message passing.
Java 21+. Java 21 language and above.
Component annotations. Declarative wiring.
Effect API. Value plus side effects.
Local dev console. Topology, traces, state.
Testing framework. Unit and integration kits.
Maven plugin. Scaffolding, build, packaging.
SBT plugin. Same for Scala projects.
Hot reload. Code changes without restart.
CLI. akka command for deploy, logs, exec.
IDE integration. IntelliJ and VSCode plugins.
Serialization. Protobuf, JSON, custom codecs.
Component versioning. Schema evolution rules.
Data & state
Event sourcing. Immutable event per state change.
Event journal. Durable queryable event log.
Snapshotting. Periodic state snapshots.
CQRS. Separate command and query paths.
Materialized views. Event-driven read models.
In-memory state grid. State cached across cluster.
Per-entity consistency. Commands serialized per entity.
Cross-entity consistency. Async projections.
At-least-once delivery. Retry with idempotency keys.
Exactly-once effects. Effect ledger dedupes retries.
CRDT support. Distributed Data replicated types.
Key-value entity. Non-event-sourced state.
Schema evolution. Versioned events and states.
Agentic AI
Agent runtime. Session-scoped LLM with retry.
Session memory. Durable per-session context.
Multi-agent orchestration. Teams as workflow steps.
Tool binding. MCP consumption from any component.
Tool exposure. Any component becomes an MCP tool.
Agent-to-agent (A2A). Cross-service agent RPC.
Streaming responses. Token-level to clients.
Structured outputs. Schema-validated JSON.
Function calling. Typed argument marshalling.
Model provider abstraction. Swap models per agent.
Prompt templates. Versioned reusable prompts.
Context window management. Auto compaction, summary.
RAG integration. Retrieval before generation.
Embeddings. Vector generation and storage.
Semantic search. Vector, keyword, hybrid.
Trace propagation. End-to-end agent trace.
Model support
Anthropic Claude. Claude 5, 4.8, 4.7, Haiku 4.5.
OpenAI. GPT-4 family and o-series.
Google Gemini. Via AI Studio API.
Google Vertex AI. Enterprise-hosted variant.
Fireworks. Hosted open-weight models.
AWS Bedrock. AWS multi-model service.
Azure OpenAI. Azure-hosted GPT.
Self-hosted open-weight. Gemma, Llama, Qwen, Mistral.
Local inference. Ollama and vLLM endpoints.
Custom endpoints. Bring-your-own HTTP or gRPC.
LoRA adapter serving. Task adapters over a base.
Model routing. Per-request policy selection.
Inference routing
AI gateway router. Routing decisions a proxy applies.
Agent-blind routing. Agents never name a model.
Two-stage pipeline. Declared rules first, meaning second.
Rulesets. Named rules enforced as one set.
Six predicates. Equals, prefix, suffix, contains, in, exists.
No regex predicate. Closed set, no request-path stall.
Protocol-agnostic attributes. One rule across OpenAI and Anthropic.
Prompt-text matching. Whole prompt, or last user turn.
Header matching. Any request header, lower-cased.
Closed destination set. Undeclared destinations refused at write.
Immutable revisions. A revision number is never reused.
Conditional writes. If-Match and ETag on every write.
Ruleset history. Every change, revision, and author.
Rule dry run. Try a request before shipping the rule.
Classifier registry. Model providers a rule may consult.
Immutable classifiers. A changed prompt is a new id.
Classifier deferral. A rule asks a model to decide.
Conversation pinning. Later turns follow the first model.
Embedding-based matching. Vector arithmetic, no model call.
Zero-token decisions. Semantic routing costs no tokens.
Discovery mode. Traffic observed, nothing routed.
Routing mode. A manifest routes unmatched requests.
Embedder identity gate. Manifest refused under a changed embedder.
Frozen route header. Header name hashed into the manifest.
Tag bound once. A tag never points elsewhere later.
Evidence log. Input, time, and rule version per decision.
Exemplar privacy gate. Text withheld below a sighting threshold.
Retention opt-out. Drop the stage, retain no text.
Capture stages. Summary line, local file, object storage.
Gap markers. Exchanges never recorded, named as such.
Content never logged. Bodies leave only through a recorder.
Model serving
vLLM serving. Open-weight models on your own GPUs.
Model descriptor. One file deploys a model.
Descriptor validation. Checked before anything is deployed.
Descriptor export. Running deployments written back to file.
Accelerator classes. GPU hardware carved into placement capacity.
Dedicated or shared. Cards held by one model or several.
VRAM budget. The share of a card per model.
GPU inventory. Kind, memory, free count, per node.
Cards per node. A multi-card model needs one machine.
Portable accelerator names. One descriptor across unlike regions.
Tensor parallelism. One model split across cards.
Pipeline parallelism. One model split across stages.
Data parallelism. Replicated weights across cards.
Expert parallelism. Mixture-of-experts sharded across cards.
Prefill/decode disaggregation. Separate fleets for each generation phase.
Continuous batching. Sequence and batched-token limits per deployment.
Chunked prefill. Long prompts split across scheduler steps.
Quantization. Weight precision set per deployment.
KV-cache dtype. Cache precision set per deployment.
CPU offload. Host memory holds weights past VRAM.
Swap space. Host memory backs the KV cache.
LoRA adapter serving. Task adapters over one shared base.
Adapter by request. A base:adapter name selects the adapter.
Adapter discovery. Found under an object-storage prefix.
Queue-depth autoscale. Replicas follow requests waiting.
KV-cache autoscale. Replicas follow cache pressure.
Rolling model updates. Surge, unavailable, and min-ready bounds.
Inference routes. Which models answer on which hostname.
Model proxy. A local port onto a deployment.
Tool-calling parser. Parser and chat template per model.
Max model length. Context ceiling set per deployment.
Hugging Face source. Token reference and weight cache volume.
Akka Optimize
Observe. Measure work agents do on live traffic.
Route. Send each task to the fitting model.
Train. Specialize models on your own data.
Serve. Run in your environment, frontier fallback.
Targeting engine. Scores narrowness, volume, spend.
Deterministic grader. Structured prompts (SQL, JSON).
LLM judge. Scores unstructured prompts.
Grades external harnesses. Copilot, Claude Code, Cursor.
Governance signal. Scores drive audit trail.
Training signal. Scores drive reward model.
Promotion signal. Scores drive promotion.
Open-weight base tuning. 4–12B base with RL.
LoRA adapter. ~50MB weights you own.
In-VPC training. GPUs in your cloud account.
Air-gapped training. Supported where required.
Fireworks training. Alternative to in-VPC.
Quality gate. Quality ≥ baseline before promotion.
Token gate. Token drop ≥ target before promotion.
Shadow staging. Candidate runs alongside incumbent.
Canary promotion. Gradual rollout by traffic share.
Fleet promotion. Full rollout after gates clear.
Before/after benchmark. Saved to audit record.
Post-promotion grading. Catches drift live.
Auto-retrain on drift. Fresh candidate to staging.
Auto-revert on slip. Return to proven model.
Improvement policy. Bounds tuning scope and targets.
Interaction log. Every decision written durably.
AI Gateway. OpenAI-compatible ingress.
Tier select. Per-request model tier.
Adapter hot-swap. Live swap without restart.
Response cache. Gateway-level caching.
Gateway rollback. Canary, shadow, rollback controls.
Frontier fallback. GPT-5.5, Opus 4.8 as teacher.
Trace store. Agent traces retained.
Scrub & curate. PII and secret redaction.
Judges & evals. Reward-model pipeline.
RL environments. Sandbox training suites.
Synthetic data. Teacher and RLVR generation.
SFT · DPO · RL. Three training regimes.
Elastic inference. Base plus adapters, auto-scale.
BYO-agent plugin. Endpoint config, trace capture.
Log forwarder. Prod logs to gateway.
Repo & CI connection. Codebase, sandbox deploys.
Budget-owner console. Spend, savings, cost per task.
AI-owner console. Pipeline, promoted-model quality.
System-owner console. Throughput, ladder, health.
90-day assessment. Install, Observe, Witness.
Reinforcement learning. Runs trained through TRL and verl.
Engine by GPU count. RL engine picked from cards given.
Smoke check. A real slice trained before full spend.
Smoke verdicts. Six named outcomes, ungraded among them.
Pause and resume. A run stopped and started again.
Durable checkpoints. A candidate outlives the job that made it.
Per-run cost budget. A ceiling set before a run starts.
Training-job logs. Job logs read without cluster access.
Content-hashed datasets. The same bytes are one dataset.
Separate evaluation set. Training and scoring data named apart.
Akka Verify
Batch evaluations. Scheduled offline runs.
Online evaluations. Continuous live scoring.
Guardrails. Input and output policy enforcement.
Policy records. Machine-readable per decision.
Evidence export. Regulator-ready audit bundles.
Per-decision audit trail. Inputs, outputs, policy.
SAFR mapping. Safeguards for agentic finance.
Regulatory pattern library. 49 regs, 227 controls.
Explainability records. Reasoning trace per output.
Drift detection. Behavior-change alerts.
PII detection. Auto-redaction hooks.
Prompt-injection defence. Layered input filters.
Model comparison. A/B judging harness.
Red-team harness. Adversarial test suite.
Tool-permission metric. Tool names against an allow list.
Tool-correctness metric. Tool names against expected tools.
Argument-correctness metric. Tool arguments against expected arguments.
Latency budget. Recorded latency against a duration.
Token budget. Input plus output tokens against a number.
Model-call budget. Model calls against a limit.
Retrieval metrics. Recall, precision, reciprocal rank, NDCG@K.
Faithfulness metrics. Reply and citations against retrieved passages.
Plan metrics. Plan quality, adherence, and step efficiency.
Bring-your-own judge. You supply the model, we supply the prompt.
Rubric override. Your rubric beats the built-in id.
Judge-call cap. A ceiling on judge calls per run.
Checkpoint evaluation. A trained model scored before deployment.
Tool-call evaluation. The requests an agent made, scored.
Red teaming
Adversarial campaigns. Attacker-shaped prompts against your own system.
AILuminate v1.1 hazards. Results grouped by the published taxonomy.
OWASP LLM Top 10. The second classification each result carries.
Three-verdict scoring. Broke through, held, or inconclusive.
Cost-ordered tiers. Deterministic, then heuristic, then model judge.
Secret exfiltration. The reply quotes a withheld marker.
Prompt-injection detection. The injected marker read back out.
PII leak detection. Luhn-valid cards, SSN patterns, configured identifiers.
Jailbreak scoring. Refusal absent, task-fulfilment language present.
Guardrail scoring. Reply checked against a configured policy.
Refusal consistency. An early refusal reversed later on.
Agentic jailbreak judge. A model reads the undecided exchanges.
Attack-goal judge. Decides when an attacker reached its goal.
Static attack wrappers. Base64, ROT13, multilingual, splitting, role play.
Adaptive attackers. PAIR, Crescendo, and Tree of Attacks.
Versioned corpora. Date-versioned sets, each with attribution.
Bring-your-own judge. You supply the model, we supply the prompt.
Breaks and holds. Two counts, never a single rate.
Akka Specify
Dev Spec. Goals, requirements, knowledge sources.
Eval Matrix. Risk, operational, improvement envelopes.
Goals & intent. Purpose, success metrics, autonomy.
Functional requirements. What the system must do.
Knowledge sources. What the system must know.
Risk envelope. Safeguards it must not cross.
Operational envelope. SLA, regions, latency, cost.
Improvement policy. What it should optimize toward.
Versioned specs. Every spec tracked with history.
Build track. PM, Dev, ML/Data Eng, Analyst.
Govern track. Risk, Security, Compliance, SRE, FinOps.
Continuous build. Spec change regenerates the system.
Continuous governance. Eval Matrix enforced each build.
Guaranteed outcome. System is the deliverable.
99.9999% availability. UP guarantee.
Governance enforced. SAFE guarantee.
Improves in production. SMART guarantee.
Weeks, not quarters. Delivery timeline.
Fixed price. One number, build plus run plus govern.
Delivered and operated. Not handed off.
Isolated VPC install. Optionally in your VPC and region.
Private data link. Governed least-privilege access.
Data under your control. No data leaves customer plane.
24/7 SRE operations. Akka SREs run avail, safety, latency.
Includes the Akka platform. In the fixed price.
Includes all infrastructure. In the fixed price.
Includes all AI tokens. In the fixed price.
Includes model training. In the fixed price.
Includes delivery. Implementation in the fixed price.
Includes 24/7 SRE. Operations in the fixed price.
Integration & protocols
HTTP endpoints. REST, streaming, SSE.
gRPC endpoints. Bidirectional protobuf streams.
MCP tool binding. Consume MCP from components.
MCP tool exposure. Expose components as MCP.
Agent-to-agent (A2A). Cross-service agent RPC.
Brokerless messaging. Component pubsub in-process.
Kafka bridge. Consume and produce topics.
Pulsar bridge. Pulsar equivalent.
CDC ingestion. Debezium-style change streams.
Server-sent events. SSE for streaming clients.
WebSocket. Long-lived client connections.
Deployment & operations
Regions. Multi-region deployment.
Geo-pinning. Traffic and residency by region.
Rolling updates. Zero-downtime rollout.
Blue/green. Parallel-version staging.
CI/CD hooks. Actions, GitLab, Jenkins.
Secrets management. Encrypted, runtime-injected.
Network policies. Service-to-service rules.
CVE patching. Managed cadence, opt-in windows.
Deploy from CLI. akka deploy per project.
Deploy from console. Click-through from AAO.
Deploy from API. Programmatic control.
Project isolation. Per-project quota, fences.
Environment isolation. Dev, stage, prod.
Resource limits. CPU, memory, concurrency caps.
Identity & access
RBAC. Four built-in roles.
OIDC. Federated login via any provider.
SAML. Enterprise SSO integration.
Project membership. User-to-project mapping.
Environment isolation. Grants per environment.
API tokens. Long-lived and short-lived.
Service accounts. Non-human identities for CI.
MFA. Multi-factor for console access.
Role-change audit log. Every grant recorded.
Session management. Lifecycle and revocation.
Security & compliance
In-VPC deploy. Runtime in your VPC.
In-account deploy. Runs in your cloud account.
Air-gapped. No internet dependency.
Sovereign. Customer controls all data plane.
TLS 1.3. Encryption in transit everywhere.
AES-256 at rest. Encryption for persisted state.
Key rotation. Managed or bring-your-own KMS.
Audit logging. Control-plane actions logged.
SOC 2 Type II. Annual audit, NDA report.
ISO 27001. Annual certification.
GDPR. DPA and subprocessor list published.
HIPAA readiness. BAA available on request.
DPIA support. Assessment inputs supplied.
Subprocessor list. Public, notified changes.
Observability
Distributed tracing. Per-request across services.
Agent trace. Per-invocation LLM trace.
Workflow trace. Step-by-step execution trace.
Metrics export. Prometheus and OTLP.
Event-journal query. Per-entity time-series.
Component health. Live status per component.
Log aggregation. Structured JSON with context.
Grafana integration. Dashboards and alerts.
Datadog integration. APM, metrics, logs.
Prometheus integration. Native scrape endpoint.
OpenTelemetry. Traces, metrics, logs on OTel.
Structured logs. Schema-aware format.
