Engineering deep-dives, product announcements, and API changelog.
Fine-tune moderation behavior without retraining models. The HushPolicy YAML DSL lets you raise thresholds for specific contexts, retrieve semantically similar examples for edge-case calibration, and inject custom prompts into the Tier 3 LLM judge.
Read moreEvery moderation request needs to be logged, usage metered, and potentially replayed. In this post, we walk through why in-process goroutines weren't enough, and how we built a durable, idempotent SQS consumer with graceful drain on shutdown.
Read moreMost moderation APIs tell you what was blocked. HushSafe tells you why, and more importantly — what would need to change for the decision to flip. Here's the design behind our counterfactual explanation endpoint.
Read moreWe replaced Detoxify with a direct ONNX export of unitary/multilingual-toxic-xlm-roberta to get real per-language scores, eliminate fabricated constants, and cut inference latency by 40% through INT8 quantization.
Read more