Bring your own data
Drop in CSV, TSV, text or Excel worksheets. DAS detects headers, handles worksheets, reads timestamp columns and lets you pick one or many numeric series to analyse.
Private by default
The file stays in your browser. That makes it useful for production extracts, customer exports, incident timelines and sensitive operational data that should not be uploaded elsewhere.
Explainable findings
Every breach gets a 0-100 score, a short narrative, the trigger window, the worst point and the weighted reasons behind the score.
Useful outputs
Export the results as CSV or JSON, or take the chart as a PNG. The output is designed for incident notes, handovers, reports and follow-up analysis.
DAS can choose the detector for you.
Real data rarely arrives with a label saying “this is seasonal” or “this is a stuck sensor”. Press Identify and DAS studies the selected columns, scores the likely detector families, tunes the winning configuration and runs the analysis immediately.
It looks for the shape of the problem: timestamp gaps, frozen runs, repeated flapping, volatility bursts, daily or weekly rhythm, lasting level shifts, sudden shocks, one-sided excursions and peer outliers across comparable columns.
The recommendation is not a black box. DAS shows confidence, the signals it found and the nearest alternatives, so you can accept the suggestion, switch to a runner-up, or fine-tune the settings yourself.
DAS is built for different shapes of failure, not one generic anomaly button. Each detector is configurable, documented inside the app and backed by a live example graph.
Static threshold
Flags values above, below, inside or outside a fixed boundary. Best when there is a known bad limit: SLA breach, temperature ceiling, queue depth, disk usage or error rate.
Rolling baseline
Compares a recent window with a longer local baseline. Useful when the absolute value is not wrong, but it has shifted sharply from what was normal a moment ago.
Sustained degradation
Ignores a lone blip and fires only after a run of bad samples. Good for slow regressions, persistent latency, throughput drops and anything where duration matters.
Lag comparison
Compares the current window with the same-length window earlier in the data. Use it for period-over-period drift when “worse than before” is the question.
Seasonal comparison
Judges a point against the same offset in prior cycles. It suits daily and weekly rhythms where high for 3am is very different from high for 3pm.
Rate of change
Finds sudden point-to-point shocks. The value may still be within a safe range, but the speed of change is itself the warning.
Volatility burst
Detects a metric becoming erratic by comparing recent dispersion with normal dispersion. Useful for jitter, thrashing, unstable control loops and noisy feeds.
Flapping
Counts repeated crossings around a boundary. Built for chattering health checks, unstable services and alerts that repeatedly fail and recover.
Stuck metric
Catches long runs of identical or near-identical values. A sensor, exporter or data feed can look healthy while silently freezing in place.
Missing data
Looks at the timestamp cadence and flags gaps. Sometimes the incident is not a bad value; it is the absence of a value at all.
Cross-series outlier
Compares one column with its same-time peers. Ideal for fleets: one host, branch, sensor, site or service instance drifting away from the pack.
Model baseline
Builds an expected-value baseline from time context and recent history, then flags residuals. Use it when “normal” depends on context rather than a fixed number.
Binomial regression baseline
Analyses success/total proportions with their underlying counts. A drop from 98% to 90% means something very different across 10 events than across 10,000.
The score is earned, not guessed.
When a detector fires, DAS builds a score from separate components: why that detector fired, how long the breach lasted, the worst-sample magnitude, how many selected columns were affected, and an optional secondary metric you choose.
Click a highlighted point and the breakdown opens beside the chart. The top line tells you what happened in plain English; the table underneath shows the weighted evidence behind the number.
Why
How long
How far
How wide
Historical labels become probability
If your data includes an independent 0/1 incident label column, DAS can estimate the probability that each row is an incident from the feature columns you select.
It reports readiness, label counts, model quality scores and per-row contributions, so risk scoring remains auditable rather than decorative.
Useful when detection is not enough.
Detector breaches answer “what looks wrong?” Risk estimation answers a different question: “given what we have seen before, how incident-like is this row?”
That is useful for triage queues, historical post-incident review, capacity risk, customer-impact scoring and datasets where an unusual value matters only when several supporting signals line up.
Operational incidents
Take an export of latency, errors, saturation, queue depth and deploy markers; find the window where behaviour changed; export the evidence for the incident record.
Sensors and field data
Spot frozen sensors, missing readings, volatility bursts and peers that diverge from nearby devices without sending sensitive site data to a third-party service.
Product and business metrics
Analyse conversion rates, success proportions, seasonal usage, support volumes and exported KPI sheets where the signal is spread across several columns.
From spreadsheet to evidence.
DAS is live at das.dixon.cx. Open a file, press Identify, and see what the data is trying to tell you.