Latency target
p95 < 25ms added to your LLM call. We achieve this with: pre-loaded SAE weights on dedicated GPUs, batched forward passes, and INT4 quantized SAE inference (lossless per our benchmarks). Your user-facing latency stays within SLA.
Back to Watchtower
Q4 2026
Feature Firehose API
Low-latency streaming API. Your production LLM calls pass through; per-token feature activations emit on WebSocket. ClickHouse-backed dashboards. Pricing from $2 per 1M tokens monitored. The enterprise tier that funds the OSS platform.
Deployment modes
Cloud (we host, you pipe) · VPC-peered (low-latency inside AWS/GCP) · On-prem Docker (regulated industries). Watchlist updates push via MQTT so you never need to re-deploy.
What emits
Per-token: top-K feature activations with IDs, values, and semantic tags from Atlas. Aggregated: hourly feature histograms, anomaly scores, watchlist trigger counts. Everything queryable via SQL on ClickHouse.
Request early access
We prioritize researchers, educators, and safety teams who will use it publicly. Tell us what you want to build; we'll reach out when the beta opens.