Open the live demo, then click "Load sample data". No install, no sign-up, no backend.
A contact center workforce management tool: forecast interval-level contact volume with comparable methods, prove accuracy with rolling-origin backtests, convert the forecast to interval staffing through Erlang A or Erlang C, and test what-if scenarios live.
Built by Ryan Allen (WFM senior team lead; Verint tenant admin). Runs fully in the browser: no backend, no data leaves the machine.
npm install
npm run devOpen http://localhost:5173, click "Load sample data" (2 years, 3 queues, 30-minute intervals), and walk the tabs left to right. npm run build produces a static site; npm test runs the suite.
Data. Bundled sample program (two voice queues plus chat, public-sector shape: Monday peaks, benefit-cycle bumps, holiday closures, injected outages) or your own CSV with columns timestamp,queue,offered,aht, one row per interval:
timestamp,queue,offered,aht
2026-01-05T08:00,voice-support,42,415
2026-01-05T08:30,voice-support,51,402
2026-01-05T09:00,voice-support,58,398MAD-based outlier cleaning runs before any model sees the data, and the flagged points are listed, not hidden.
Forecast. Three component methods plus a custom ensemble, all hand-implemented in TypeScript:
- Seasonal moving average: trimmed, recency-weighted same-weekday average (the long-horizon benchmark the literature says is hard to beat).
- Holt-Winters: additive, weekly seasonality, grid-searched smoothing parameters.
- Dynamic harmonic regression: ridge regression on weekly and yearly Fourier terms, trend, weekday, holiday, post-holiday, and month-start features.
- Ensemble: blends the three with weights proportional to inverse WAPE raised to a power picked from a small grid, fitted per horizon bucket (1-3, 4-14, 15-28 days) on non-overlapping rolling-origin folds inside the training window and scored against raw actuals, the same way the backtest judges the final forecast. The fitted weights are shown in the UI.
Daily forecasts are spread to 30-minute intervals with recency-weighted day-of-week profiles learned from cleaned history. The daily chart shades an 80% prediction range around the ensemble, calibrated from the empirical errors of the same rolling-origin evaluation that fits the blend weights.
Accuracy. Rolling-origin backtest (8 folds, 28-day horizon, re-cleaned and re-fit per fold so nothing leaks), scored as WAPE, MAPE, and bias at interval, daily, and weekly grain. The scorecard includes an equal-weight blend row as a benchmark, so the fitted ensemble weights have to prove they beat the naive average. Sample-data result for the largest queue, daily WAPE: seasonal average 11.2%, Holt-Winters 11.3%, harmonic regression 7.9%, equal-weight blend 9.5%, ensemble 8.1%. The scorecard renders whatever the loaded data says, including when a single component beats the ensemble: on this data the harmonic regression still edges the ensemble by 0.15 points.
Staffing. Interval requirements via Erlang C or Erlang A (abandonment-aware; the birth-death solve was validated against an independent 400k-arrival Monte Carlo simulation to within 0.3pp on service level). Levers: SL target, patience, abandonment cap, shrinkage (applied as division, the correct way), occupancy cap, chat concurrency, volume and AHT deltas. Scenario B comparison shows the FTE-hour and occupancy deltas per day. A fixed-staff ("what I have") mode projects SL, ASA, and abandonment at the heads you enter instead of solving for a target, flagging the intervals that miss. An optional cost-per-hour rate prices scheduled FTE-hours.
Working with results. Forecast, scorecard, and staffing tables download as CSV. Scenario settings encode into the URL hash, so a what-if is shareable as a link. Forecasts, backtests, and staffing solves run in a Web Worker to keep the sliders responsive, and a header toggle switches between light, dark, and system theme.
Research notes with sources are in docs/research.md: Taylor 2008 on which classical methods win at which horizons, the forecast-combination evidence, Erlang A vs C, pooling math, and the accuracy-metric tradeoffs. The design rationale (stack, data model, algorithm spec) is in docs/design.md.
This release is module 1, the forecast + staffing engine. Modules 2-4 (capacity planner, queue strategy analyzer, intraday reforecast simulator) are specced in docs/design.md.
Vite, React, TypeScript, Recharts. All forecasting and queueing math lives in src/engine as dependency-free pure functions with Vitest coverage, including Erlang B/C/A against published table values.
MIT, see LICENSE. The agent-harness files under .claude/, .agents/
and .codex/ are template tooling rather than project source; some carry their
own terms, listed in NOTICE.