For Platform & MLOps teams running AI at scale
Three deployment paths. Plugin system. Autonomous healing. 30-day degradation forecast. Built for the SREs who get paged at 3am.
Sgraal already runs in production with 15ms p50, 99.95% uptime, and zero-dep edge mode for air-gapped deployments.
Hosted for most teams. Self-hosted for regulated sectors. Edge mode for air-gapped, classified, or sub-millisecond latency budgets.
sgraal.com · default
~15ms p50 latency. Full 87-module engine, fleet vaccine corpus, per-tenant calibration, Q-learning per-domain. No infra to manage.
Quickstart
from sgraal import SgraalClient client = SgraalClient("sg_live_...") result = client.preflight(memory_state, domain="fintech")See pricing →
Docker / Helm · on-prem
Run Sgraal in your VPC. Docker image + Helm chart. Same engine, same API surface. Your tenant calibration stays inside your network. Enterprise tier.
Quickstart
# Helm
helm repo add sgraal https://charts.sgraal.com
helm install sgraal sgraal/sgraal \
--set apiKey=sg_live_... \
--set sgraalConfig.path=/secrets/cfg.json
Request Helm chart →
in-process · zero-dep · sub-ms
Runs inside your Python process. Zero dependencies — no numpy, no requests. For air-gapped, classified envs, or trading desks that can't take ~15ms. Subprocess-isolated tests prove zero network leak.
Quickstart
from sgraal.edge import edge_preflight result = edge_preflight( memory_state, domain="fintech") # < 1ms p99, no networkSee the cut line →
Per-tenant isolation, registry-only loading, escalate-only security monotonicity. Install custom security rules that can downgrade USE_MEMORY → BLOCK but never the reverse.
DNS-pinned dispatch to prevent DNS rebinding. SSRF-safe URL validation. Built-in formatters for Slack, PagerDuty, OpsGenie. Retries with exponential backoff + dead-letter queue.
Call frequency trend, action-type escalation pattern, domain switching. Detect compromised or drifting agents before they cause incidents. Surfaced via /v1/agent/{id}/behavioral-profile.
Closed-loop reinforcement learning from real outcomes. Report /v1/outcome after each agent action; Sgraal adapts its policy per-domain. Tenant-isolated Q-tables.
/v1/predict/degradation returns the projected day a given memory state will cross the BLOCK threshold. Schedule refresh ahead of customer impact.
Zero-dep in-process scoring with per-type thresholds (identity, policy, semantic, preference, episodic, shared_workflow, tool_state). Sufficient for many production deployments without calling sgraal.com. For workloads requiring fleet-wide vaccine signatures or cross-tenant analytics, use the hosted or self-hosted mode.
Web UI for tuning your scoring weights, Weibull λ values, and detection thresholds. Live A/B replay against your own 30-day history. Today: edit SGRAAL_SCORING_CONFIG env var directly.
Deterministic replay of past decisions with counterfactual branching. "If T-3 we had healed memory X, would T-2 decision have been different?" Git bisect for AI cognition.
Be first to know →/v1/preflight/stream emits 23 events across 4 phases (15 scoring module completions + 6 detection layer states + invariant check + final decision). Tap into it for live dashboards or live debugging.
curl -N https://api.sgraal.com/v1/preflight/stream \ -H "Authorization: Bearer sg_live_..." \ -d @memory.json
/v1/admin/endpoints returns the full endpoint inventory; /v1/admin/dlq exposes the dead-letter queue for failed webhook deliveries. Both gated by SGRAAL_ADMIN_TOKEN.
Returns 404 (not 401) if the admin token is not set on the server — prevents endpoint enumeration.
/metrics endpoint gated by SGRAAL_METRICS_TOKEN. Exposes per-tenant decision counters, latency histograms, circuit-breaker state, fleet vaccine propagation rate.
SHA256-chained decision log per tenant. Export via /v1/audit-log/export in CEF or LEEF for Splunk, Datadog, Elastic.
15ms p50 hosted. Sub-ms p99 edge. Air-gapped self-host. Same wire format, same decision contract.