Fixed thresholds tend to either produce false alarms or miss real anomalies. In Scandock, AI judges whether something is abnormal by combining value trends, past history, and the facts in your project's memory. It's anomaly detection that doesn't depend on thresholds.
DB · Payment success
Failures over the last 30 minutes are 6x the usual level. They're concentrated on one card issuer, while successes are normal.
Successes rose along with the lunchtime order increase. The failure rate is within the usual range.
Values are at the same level as this time yesterday.
Loosen thresholds and you miss things; tighten them and false alarms pile up
Business context like "orders dip at month-end" can't be captured in a threshold
When no alerts come, you can't tell if all is well or monitoring has stopped
Values since the last judgment are compared against earlier history. AI also reads from the timestamps whether a value normally varies by time of day or day of week.
What happened before on the same monitor, and whether it turned out real or a false alarm, feeds into the judgment.
Judgments take known facts into account, like "an increase at this time of day is normal." The same false alarm doesn't repeat.
Judgments are generated with the reasoning written out first and the verdict last. This prevents mismatches like reasoning that says nothing is wrong followed by an "abnormal" verdict.
Each time it judges, AI works from the following information.
Values collected since the last judgment, plus the min, max, and average of earlier history summarized by period.
Incidents that occurred on the same monitor, and how they were resolved.
Known facts about the project, recorded by AI and by members.
When AI judges something abnormal, it assigns one of three severities: Critical, Warning, or Info.
Each judgment's result and reasoning are saved per monitor, and the judged ranges are overlaid on the chart.
While a monitor has an open incident, no new incident is created for it.
All you need is a database connection string or an account connection. Nothing to install on your servers.
Collect values
Values are collected over a read-only connection at each monitor's collection interval.
AI judges
At each judgment interval, AI weighs the trend, past alerts, and memory together.
If abnormal, investigate
When AI judges something abnormal, an investigation into the cause begins. If it turns out to be a false alarm, the incident is closed without notifying anyone.
E-commerce
Don't mistake the lunchtime rush for an anomaly
History shows that orders rise every day at lunchtime. If the increase is within the usual range, it's recorded as "nothing wrong," with the reason.
Payments
Catch a skew in failures
Because successful and failed payments are watched side by side, a rise in failures alone is caught as an anomaly even when the total looks normal.
AI monitor design
AI reads your schema and real data to design what to watch, how often to collect it, and how often to judge it.
Automatic investigation
When something looks abnormal, AI digs into the real data to trace the cause, and records its conclusion and evidence.
Project memory
Facts found in investigations and business context are recorded and used in every later judgment and investigation.
Take your hands off
threshold tuning.
The reasoning behind every judgment is kept, so you can always check later what was looked at and how it was decided.