Inventory Optimizer
Demand forecasts, reorder decisions, and supplier scorecards for any SKU panel.
Drop in a CSV or Excel of SKU sales and get back forecasts with prediction intervals, (s,S) recommendations rounded to MOQ, a ranked reorder queue, OTIF supplier scorecards with Bayesian lead-time learning, a service-level frontier, stress tests, and a tool-using chat — all calibrated against the M5 Walmart dataset.
- Reference profiles
- 5
- Workspace tabs
- 8
- Backend tests
- 188
Or load a realistic demo dataset
What it does
An ops workspace, not a directory.
Most forecasting tools dump a number on you. This app starts with an action queue — what needs ordering today, what's at risk in four weeks — and lets you drill in when you need to. Every model decision is auditable: ensemble weights, train cutoff, and per-horizon conformal coverage are first-class artifacts on every forecast.
Forecasts you can defend
Weighted ensemble of ETS, Amazon's Chronos-Bolt, and a global LightGBM model. Weights set by inverse out-of-fold loss. Intervals are split-conformal calibrated, with empirical coverage reported per horizon (h=1 and h=4) on every SKU page.
Policies that match the SKU
Stable items get (s, S); perishables get newsvendor; smooth high-volume items get EOQ or (Q, R). Service level is auto-derived from the underage/overage ratio unless you pin it. Schedules are rounded to MOQ and case-pack.
Honest data quality
Composite score over five components. The distribution-profile component asks "does this look like the kind of data we expect?" — separately from "is your data well-formed?". Auto-detects retail / e-commerce / B2B / pharma / spare-parts; you can swap.
Tour the tabs
Eight surfaces, one workflow.
Every dataset opens with the same left-rail navigation. Each tab does one thing well; they share the same per-dataset settings and AI insights.
Overview
Today's snapshot — KPI tiles, the ABC × XYZ heatmap, working-capital tile, and a heuristic + AI insights panel.
- Dataset-wide KPIs at the top: revenue, inventory value, days-of-cover.
- ABC × XYZ heatmap shows where revenue meets variability.
- Insights panel turns raw aggregates into one-line actions (heuristic, with optional Claude enrichment).
Reorder Queue
Ranked list of SKUs that need attention — sorted by score = stockout-prob × revenue-at-risk.
- Each row shows on-hand, reorder point, recommended quantity (rounded to MOQ + case-pack).
- Expedite flags fire when the projected stockout date sits inside the lead time.
- Joint-replenishment groups are highlighted so you bundle co-supplied SKUs into one PO.
Forecasts
Per-SKU forecasts with calibrated 95% intervals, click through to a deep dive.
- Header tiles segment the panel by status: order-now, at-risk, watch, healthy.
- Default sort is days-of-cover ascending, filtered to A and B class — what to do today.
- Per-SKU page: header block, ensemble breakdown (ETS + Chronos-Bolt + LightGBM), per-horizon conformal coverage, what-if sliders, audit footer.
Frontier
Cost-vs-service-level frontier, plus the newsvendor calculator for single-period decisions.
- Move the service-level slider; see the policy's recommended order quantity, safety stock, and total annual cost track in real time.
- Newsvendor calculator (Q*) for perishables: F⁻¹(Cu / (Cu + Co)) on the demand distribution.
- Recommended Q (multi-period (s,S)) and Q* (single-period newsvendor) answer different questions — both are surfaced with explainers.
Suppliers
Scorecards with OTIF, lead-time mean ± std, and a Bayesian posterior that updates from receipts.
- Lead-time variance, on-time %, in-full %, payment terms, MOQ — all per supplier.
- Posterior mean and std come from Gamma updating against actual receipt history.
- Insights panel calls out concentration risk and the slowest OTIF supplier by revenue share.
Stress test
Shock the inputs (lead-time, demand) and watch revenue at risk move.
- Outputs Δ revenue at risk, VaR 95%, CVaR 95%, count of SKUs over threshold.
- Top-impacted table shows base vs shock for stockout %, $ at risk, and recommended Q.
- Run before a forecasted disruption (port strike, viral spike) to see who's exposed.
Data quality
Composite score over five components, scored against a domain-appropriate reference profile.
- Completeness, plausibility, history depth, stationarity, distribution profile.
- Distribution profile uses a soft-penalty curve against one of five reference profiles (retail M5, e-commerce, B2B, pharma, MRO).
- Auto-detect picks the profile from your panel medians; you can swap in Settings.
Chat
A tool-using Claude agent with structured access to your dataset.
- Tools: query SKUs, get forecast, run scenario, compute reorder, compare to M5, decompose series.
- Streamed via SSE so partial answers render as the agent works.
- Cost-capped: 30 requests / minute / IP, with a daily USD budget guard.
Ready to try it?
Upload your panel, or boot a demo in one click. No login required.
Or load a realistic demo dataset