SYS.INT is the inference platform that routes, deploys, and observes every model call your product makes.
It sits between your application and 147 foundation models, scoring live latency and health for every request and picking the fastest route in under 400 microseconds. Deploy a model and the platform handles routing, edge replication, and failover without orchestration code.
Five systems, one deployment
SYS.INT covers routing, inference execution, edge deployment, observability, and model management as one platform. Each system runs independently monitored and none require separate configuration to turn on.
Deterministic routing
Every call is scored against live latency, load, and model health before it is sent anywhere. The router picks the fastest healthy path in under 400 microseconds and logs the decision for replay.
Sub-5ms inference
Requests run on pre-warmed compute pools sized per model. A typical call to a mid-size model returns in 4.2ms average and 12ms at the 99th percentile.
Edge deployment
A model pushed to the platform replicates to the nearest of 50+ edge regions automatically. Rollouts finish in under 90 seconds, with automatic rollback on a failed health check.
Observability
Every request gets a trace ID, a routing decision log, and a per-hop latency breakdown. Dashboards refresh every 2 seconds and retain 30 days of history by default.
Model management
Deploy, pin, or roll back any of 147 supported models by version tag. Canary rollouts default to 5% of traffic for 10 minutes before full promotion.
How does SYS.INT route a single inference request?
A request enters through the routing layer, gets scored against live model health telemetry, executes on the nearest healthy edge node, and returns with a trace ID attached. The full path completes in 4.2ms on average for a warm model.
- 01 / SCORE
The router evaluates every healthy model replica against current latency, queue depth, and error rate before a request is sent anywhere.
- 02 / SELECT
The platform selects the fastest available match from 147 supported models and locks that route for the life of the request.
- 03 / EXECUTE
The request executes on the nearest of 50+ edge regions, using a pre-warmed compute pool sized for that model.
- 04 / RETURN
The response returns with a trace ID and a latency breakdown attached, typically in 4.2ms average and 12ms at the 99th percentile.
What are the platform's technical specs?
SYS.INT's standard tier holds 4.2ms average latency and a 99.97% uptime SLA across 50-plus edge regions. Every figure below is the platform's own measured default, not a projection.
| Spec | Value | Notes |
|---|---|---|
| P50 latency | 4.2ms | warm model, cached route |
| P99 latency | 12ms | includes routing decision |
| Sustained throughput | 12.8k req/s | per region, standard tier |
| Edge regions | 50+ | auto-replicated on deploy |
| Supported model formats | ONNX, GGUF, safetensors, PyTorch | converted on ingest |
| Uptime SLA | 99.97% | Pro tier and above |
| Canary rollout window | 10 min @ 5% traffic | auto-promotes on pass |
Platform questions, answered
Straight answers to the questions engineering teams ask before they route production traffic through SYS.INT.
01What latency can I expect from SYS.INT?
Most requests return in 4.2ms average and 12ms at the 99th percentile, measured on a warm model with a cached route. A cold start on a newly deployed model adds roughly 200ms, and only for that first request.
02Which model formats does the platform support?
SYS.INT accepts ONNX, GGUF, safetensors, and PyTorch checkpoints, converting each to its internal serving format on ingest. No manual conversion step is required before deploy.
03How does SYS.INT pick which model instance handles a request?
The router scores every healthy replica on live latency, queue depth, and error rate, then sends the request to the best match. That scoring pass completes in under 400 microseconds, before the request leaves the routing layer.
04What happens if an edge region goes down?
Health checks run every 2 seconds per node. A failed node is pulled from rotation within one check cycle, and in-flight requests reroute to the next-nearest healthy region automatically.
05Can I roll back a model deployment?
Yes. Every deploy is tagged by version, and rolling back to a previous tag takes effect within the next routing cycle, typically under 5 seconds. Canary rollouts also default to 5% of traffic for 10 minutes before full promotion, so a bad deploy only ever reaches a small slice of traffic first.
Ready to route your first model?
Deploy a model and point traffic at SYS.INT to get sub-5ms routing across 50+ edge regions without writing orchestration code.
Request a Demo