{"slug": "91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp", "title": "91% Cheaper, 14.8% Faster Alert Triage with TypeSafe AI & Swamp", "summary": "Swamp Club reported that adding TypeSafe AI's System One model, version jev-1.13.0, as a first-pass alert screener cut triage costs by 91% and improved speed by 14.8%, with each TypeSafe screen costing about 99.9% less than a full Claude triage call. TypeSafe returns a probability on whether an alert indicates a real, actionable production problem, and Swamp compares that score against a threshold to either dismiss the alert or escalate it to Claude for classification, correlation, and remediation suggestions. One Claude triage call costs roughly the same as 1,365 TypeSafe screens, and Swamp retains the score, threshold, and reason for each decision so humans can review the branch taken.", "body_md": "TypeSafe AI is incredible at screening alerts.\n\n## How we got here\n\nLet the AI triage the issues, so your beautiful humans can keep sleeping until they're actually needed. I'd rather they spent their attention fixing the things that matter than reading another notification about something that doesn't need handling. Let the machine do the first pass. Give the human something worth interrupting them for.\n\nThis is how alert triage works at Swamp Club. AI does the first pass, the workflow handles the outcome, and our humans get involved when they're needed. If any part of this fails, page the human.\n\nSpeed to ping is super important for mean time to resolution, so [TypeSafe AI](https://typesafe.ai/?ref=blog.watson-labs.co.uk) now makes the first decision on firing alerts: **does this need more attention?** Swamp passes it the alert and the incident context, compares the returned probability with our threshold, and either records a dismissal or sends the event to Claude.\n\nClaude still does the richer work: classify the incident, correlate it with what we already know, explain the impact, and suggest what to do next. It just has a much more sensible queue in front of it.\n\nTypeSafe's first-pass screening costs **about 99.9% less** than our full Claude triage.\n\n## TypeSafe makes the first decision\n\nTypeSafe's System One model, `jev-1.13.0` in this configuration, returns a probability for a structured question. The workflow gets a number it can compare directly, rather than a paragraph it has to interpret.\n\nOur screening question is:\n\nDoes this monitoring alert indicate a real, actionable production problem that an on-call operator should triage now?\n\nWe define the positive case as a service failing, degrading, or measurably at risk. The negative case includes test alerts, self-recovering blips, routine deploy or scaling noise, and **restatements of an active incident already being handled**. Knowing that we're already dealing with something is part of deciding whether another alert deserves attention.\n\nThe screen sees **what's arriving now alongside what we're already handling**. We give Jev the last 20 signals and their status to help determine whether the latest one matters.\n\n## The economics of the extra question\n\n**One Claude triage call costs roughly the same as 1,365 TypeSafe screens.** Screening pays for itself if it avoids just one comparable Claude call in 1,365.\n\nEscalated alerts pick up a tiny screening charge; dismissed alerts avoid nearly all the old inference cost. But cheap is not the same as correct, so we retain the score, threshold, and reason for review.\n\n## Screen before triage\n\nI've argued before that [the pipeline is the context engine](https://blog.watson-labs.co.uk/correct-relevant-concise-in-that-order/). This is that argument with an actual decision model in it. TypeSafe sees the alert and the incident context, returns a score, and Swamp uses it to decide what happens next. Claude gets the work that passes the gate. The lifecycle code handles the things we already know how to do.\n\nWe can follow the whole chain: the incoming alert, the incident context, TypeSafe's score, the branch Swamp took, and the outcome. If the decision looks wrong, the record is there. If the workflow breaks, the failure path is there. Those are operational properties of the system we now run.\n\nTypeSafe's job is small enough to describe in a sentence and cheap enough to put in front of every firing alert. Claude's attention now goes through that decision first.\n\nWe stopped giving every alert the same expensive treatment. TypeSafe makes the first judgement. **Swamp is what makes that judgement useful:** it supplies the context, routes the outcome, records the decision, and pages a human if anything fails. Without Swamp, you have a score. With it, you have an operational system.", "url": "https://wpnews.pro/news/91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp", "canonical_source": "https://blog.watson-labs.co.uk/typesafe-ai-alert-fatigue/", "published_at": "2026-09-18 05:07:16+00:00", "updated_at": "2026-09-18 05:25:24.292369+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-products", "artificial-intelligence"], "entities": ["TypeSafe AI", "Swamp Club", "Claude", "System One", "jev-1.13.0"], "alternates": {"html": "https://wpnews.pro/news/91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp", "markdown": "https://wpnews.pro/news/91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp.md", "text": "https://wpnews.pro/news/91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp.txt", "jsonld": "https://wpnews.pro/news/91-cheaper-14-8-faster-alert-triage-with-typesafe-ai-swamp.jsonld"}}