Allocate scarce inventory to open orders and build full truckloads.
Website · Run locally · Report an issue · MIT license
When stock is short and trucks are finite, Fillrate answers three questions for a single depot shipping on 53 ft trailers: which order lines get the stock, how stops group into delivery areas, and how many full trucks it takes. Every plan is checked by an independent validator and explains why each line did or didn't ship.
- Allocate available inventory to order lines piece by piece: earliest order date first, then highest value per piece. Lines can be partly filled.
- Cluster the stops with k-means. The k explorer compares k values and seeds so you can check a grouping is stable.
- Build loads in each cluster with PyVRP: linear feet as capacity, open routes (no return leg), at most 500 miles per drive, optional delivery time windows. Fewest trucks first, then fewest miles.
- Validate and compare. Revenue, trailer fill, truck count and miles per run, cluster and shipment, with stops counted once even when an oversize stop is split across trucks; side-by-side comparison across runs and sweeps.
Distances are straight-line miles × a circuity factor (default 1.2). Roads never feed the optimizer; you can draw a shipment's road path from the public Valhalla server for display (routing © Valhalla / FOSSGIS, data © OpenStreetMap contributors).
 |
 |
| Inputs. Where the stops are, how much stock covers demand, and the open orders. | Shipments. Every truck with its trailer fill, stops, miles and value. |
 |
 |
| Timeline. Step through a shipment's drive, stop by stop, on its road path. | Unshipped. Every line that didn't ship, why, and the evidence. |
- Import orders and stock as CSV, GeoJSON points or scenario JSON, or open a bundled example in one click. Once a scenario is open, the import form folds away until you need it. Addresses geocode with the free US Census geocoder, with a labeled ZIP fallback; fix coordinates by hand on a map.
- Versioned scenarios. Edit quantities, dates and coordinates in place; each save is a new version you can branch from.
- Preflight checks flag missing coordinates, unreachable stops and oversize stops before the solver runs, each set to block or warn.
- Run inspection. Cluster map, per-shipment list and timeline, unshipped lines with reasons, stock use and full provenance.
- Experiments. Sweep k, seeds and settings across runs and rank the results.
- Exports. Printable shipment sheets, JSON, CSV, GeoJSON, and a Python bundle that replays a run with PyVRP.
Local Fillrate is the full tool: single user, no account, scenarios and runs saved in SQLite. Keep real data here.
Requires Node 24, Bun, and uv with Python 3.13:
bun install
(cd services/optimizer && uv sync)
bun run dev # terminal 1: http://localhost:3000
bun run optimizer # terminal 2: the Python solver service (127.0.0.1:8000)Open http://localhost:3000/scenarios and click Open lesson, 2,000 orders to try it.
bun run build:static exports the landing page, privacy page and /playground as static files in apps/web/out, for any static host. fillrate.blasingame.dev is this export, built and deployed by Cloudflare Pages on every push to main. The playground shows the bundled sample plan only; nothing runs on a server.
The playground sample is a bundled solver result, so viewing it does not run the optimizer. After changing the example, optimizer, or its dependency lock, regenerate it with cd services/optimizer && uv run python ../../examples/generate-playground.py. Tests reject stale sample results.
Fillrate is a planning and comparison tool, not dispatch, tracking or navigation. Linear feet are a one-dimensional capacity model, not a packing or compliance check. Allocation is an experiment against an inventory snapshot, not a reservation. Shipment counts and miles are the solver's best found values, not proven optimal.
Built with Next.js, SQLite, Python and PyVRP. Contributors and coding agents, see AGENTS.md.