Updates
Follow product changes, upcoming plans, and announcements around Howzer.
The first model download drops from around 20 GB to roughly 11.7 GB, every model is fixed to one exact version, and when something does go wrong the installer names the real cause. Existing installations upgrade in place.
Read moreMixed feedback is now read correctly 91 percent of the time instead of 40, wholehearted praise no longer lands in the crisis pile, and an important sentence buried in a long message is no longer missed.
Read moreSeven failures that never announced themselves are fixed. Reports are fully German and print-ready, the installation is around 6.3 GB smaller, and an analysis that cannot do its job says so instead of quietly delivering something worse.
Read moreThe response now carries the full picture instead of a summary, purely as an addition. Plus 17 percent faster answers and far more accurate triage of long messages.
Read moreSentiment now runs on a German model we built ourselves, scheduled reports land in your inbox automatically, and busy days stay calm.
Read moreReliability as a feature: honest results under load, suggested replies with no invented facts, nothing lost quietly, and a first installation that runs through without hand-fixing anything.
Read moreOur biggest release. Every analysis runs on models trained for German customer feedback, and around them sits the full picture: where someone stands, where they are heading, and what to do about it.
Read moreSpreadsheets import exactly as they were exported. Over 50 channel labels fold into nine. Postcode, city, reference numbers and urgency are read straight out of the running text.
Read moreSeven rules watch the live stream: the single customer about to leave, and the pattern across many messages that no individual message reveals.
Read moreA hardening release with nothing new to learn: seventeen rough edges filed off across text handling, storage, monitoring and Windows installations.
Read moreThe German-optimised pipeline is available: broader risk detection, more precise root cause analysis, and deployment tooling that survives a real data centre.
Read moreA detailed walkthrough of the pipeline, from analysing a message to drafting a reply. Written for technical and non-technical readers alike.
Read moreThe first version looked for words like cancel and lawyer. Today risk is read from the shape of a conversation rather than from its vocabulary.
Read moreWe tried the obvious approach first: one large multilingual model. Good benchmark numbers, disappointing results on real messages. Here is what it kept missing.
Read moreNot a translation layer: English messages are read by models calibrated for English, and every message reaches the right pipeline automatically.
Read moreYour messages live in mailboxes, chats and ticketing systems. Getting them into Howzer should not cost a firewall exception, or a copy of your data sitting at a vendor.
Read moreRedaction is not a setting but the first step, before any model sees a single word. The quality of the analysis is unaffected.
Read moreThe preview shows two things: risk while it is still forming rather than in tomorrow morning summary, and the path from a scored message to an approved answer.
Read moreWhat it really takes to run Howzer inside your own network: no traffic leaving, connectors behind the firewall, your own keys, and a working setup in four days.
Read morePages come from the nearest location instead of a single server, every change gets its own preview link, and downloads download instead of opening in a tab.
Read moreA clear line between what exists today and what is coming: self-hosted is available, the hosted option is announced, and neither pretends to be the other.
Read moreService teams are not indifferent, they are outnumbered. Our answer is a system that reads every message properly and hands the human a clear picture instead of answering for them.
Read more