selected work
DevTools2025

Watchtower

Uptime monitoring with incident-based detection and a cost-aware multi-model AI layer that cut operating cost ~60%.

Sole engineer — design, build, ship
  • Node.js
  • Background workers
  • SMTP
  • Multi-model AI
the problem

Naive monitors alert on every failed check and bury real incidents in noise. And once you bolt AI features onto a monitoring system, cost explodes if every task hits a frontier model — most of them don't need one.

Watchtower had to detect genuine incidents (not blips), alert reliably on both downtime and recovery, and keep the AI layer cheap without sacrificing quality.

approach

Incident-based monitoring engine

Automated HTTP health checks run through secure background workers. Detection is incident-based rather than per-check, so a single flaky response doesn't page anyone — and email alerts fire on both downtime and recovery so on-call always knows the current state.

Cost-aware multi-model AI routing

The AI layer routes each task by complexity: frontier models (Claude 4.8) are reserved for hard reasoning, while routine parsing is delegated to cheaper models (Claude 4.6, GPT-4o-mini). A memory/caching layer eliminates redundant calls entirely.

Together these cut AI operating cost by roughly 60% with no measurable drop in output quality.

architecture
  1. 1
    Scheduler

    enqueues health-check jobs

  2. 2
    Secure background workers

    run HTTP checks

  3. 3
    Incident engine

    distinguishes blips from real downtime

  4. 4
    SMTP alerts

    downtime + recovery notifications

  5. 5
    AI router

    complexity-based model tier + memory/cache

outcomes
−60%
AI operating cost
Incident-based
Detection
Down + recovery
Alerts
Secure, async
Workers

Want something like this shipped for your product?