One product, four systems, four different weights — this project measures exactly what that costs per year, and shows why no single-system fix makes it stop.
Live: https://dimensions.lailarallc.com
Inconsistent product weights and dimensions across ERP, WMS, GDSN, and DTC systems silently bleed margin through freight misclassification, parcel reweigh back-bills, and compliance chargebacks. This project quantifies that cost for a 50-SKU specialty food portfolio and presents the finding as an interactive, executive-readable web story.
Under the hood it is a full data pipeline:
- Generate — synthetic extracts of the same 50-SKU catalog as four systems see it: NetSuite (ERP), the warehouse WMS, GDSN (retail data sync), and Shopify (DTC)
- Reconcile — dbt models establish the WMS physical measurement as the measurement of record and compute each system's divergence from it
- Cost — freight class, dimensional weight, and chargeback exposure are computed per SKU from calibratable rate tables (
config/cost_params.yml, every parameter source-cited) - Present — results export to JSON and render as a React narrative with interactive "fix this system" toggles
The four steps run as a Dagster asset graph: generate_source_extracts → load_raw → dbt_build → export_frontend_json.
Each cost lane is billed by a different party, so no single report ever shows the total. For the hero SKU (CHP-AS-002, Roasted Garlic Marinara), which carries four conflicting physical representations across four systems:
- LTL freight reclassification. GDSN publishes inflated dimensions yielding density 37.98 lb/ft³ and freight class 55 instead of the correct class 50. Cost: $0.39/case × 10,851 cases/yr = $4,231.89/yr.
- Parcel reweigh back-billing. Shopify lists ship weight as 1.00 lb (unit net weight). The actual parcel weighs 2.05 lb, billable at 3 lb. Cost: $1.97/shipment × 200 orders/yr = $394/yr.
- Compliance chargebacks. Published dimensions do not match physical measurement. Cost: $200/event × 1.2 expected events/yr = $240/yr, risk-adjusted for the ~40% of divergent SKUs that actually incur chargebacks.
Total annual cost for one SKU: $4,865.89. Across the 50-SKU portfolio: $208,310.87.
LTL is billed per hundredweight, so the reclassification penalty falls on annual tonnage shipped — how cases stack on a pallet cancels out of the arithmetic entirely. The driver is therefore priced per case against an annual case volume, derived from disclosed catalog figures rather than an asserted pallet count (see config/cost_params.yml).
The deeper finding is the localization paradox: fixing retail data (GDSN → physical) clears the LTL reclassification but leaves the DTC parcel leak. Fixing DTC data (Shopify → actual weight) clears parcel back-billing but leaves the LTL overcharge. No single-system toggle clears both — only a governed measurement of record does. The live frontend lets you flip each fix and watch the cost move.
Built on the Cinderhaven synthetic dataset — a ~$25M specialty food brand, 50 SKUs across 5 product lines and 6 contracted retailers. The data is synthetic; the methodology, rate tables, and deliverables are real.
Canonical baseline: 50 SKUs · 5 product lines (AS·PS·SC·DG·SB) · 6 retailers (Walmart·Costco·Whole Foods·Sprouts·Kroger·Regional Group) · 10 channels (6 retail + UNFI·KeHE·DPI + DTC). Physical-attribute fields are owned by this project; the companion Product Data Health Audit owns structural completeness.
The frontend reads committed JSON exports in frontend/src/data/, so it runs standalone:
cd frontend
npm install
npm run dev
Python dependencies are pinned in requirements.txt (dbt-postgres, dagster, psycopg2, pyyaml, pytest):
python -m pip install -r requirements.txt
Postgres connection is configured by environment variables CINDERHAVEN_DB_HOST / CINDERHAVEN_DB_PORT / CINDERHAVEN_DB_USER / CINDERHAVEN_DB_PASSWORD / CINDERHAVEN_DB_NAME (defaults: localhost:5432, user postgres, database cinderhaven).
From the repo root:
export PYTHONPATH=.
dagster dev -f dagster/definitions.py
Then materialize all four assets from the Dagster UI. The steps can also run individually, e.g. python -m data_gen.generate_dimension_mess to regenerate the source CSVs into data/generated/.
# Python tests (cost math, data generation, export, E2E reconciliation)
python -m pytest tests/ -v
# Frontend tests (Vitest)
cd frontend && npm test
To validate a client's item master and compute its physical attributes locally (no database, no deploy), install the shared scaffold and run client mode:
python -m pip install -e ../engagement-template/lib
python client_mode.py --config engagement.yml --input client-data/item_master.csv \
--out client-output [--final]
It reads CSV/XLSX tolerantly (SKU kept as text), runs a preflight that emits a
branded Data Readiness Report if a required case dimension/weight column is
missing, then computes cube, density, and NMFC freight class per SKU into a
branded, provenance-footed, DRAFT-watermarked report in client-output/
(gitignored). Required fields, mapping, and scope are in
INPUT-SPEC.md. engagement.demo.yml is a safe-to-deploy
example (demo: true); a real engagement.yml is runtime-only and never
deploys (scripts/engagement_guard.py enforces this). The four-system
divergence and dollar-cost model below is the dbt pipeline and is unchanged.
Pushing to main deploys the site. The Cloudflare Pages project is
connected to this GitHub repository, so it builds and publishes
dimensions.lailarallc.com on every push. No manual step is required, and no
credentials are needed locally.
frontend/package.json also carries a deploy script
(wrangler pages deploy dist) for a manual direct upload. It requires an
interactive terminal. Wrangler authenticates through a browser OAuth flow,
and its stored token expires; once it has, the script fails in any
non-interactive shell — CI, a script, or an agent session — with:
Not logged in. Your auth token has expired and could not be refreshed,
and the environment is non-interactive.
To use it, run wrangler login in a real terminal first (or set
CLOUDFLARE_API_TOKEN in the environment). Because the Git integration
already covers normal deploys, reach for this only to publish a build that
is not on main.
- Pipeline: Python 3.13, dbt-core, Dagster, PostgreSQL
- Frontend: React 19, TypeScript 5.7, Vite 6
- Deployment: Cloudflare Pages (frontend), Fly.io (pipeline)
- Design: Lailara design system (Playfair Display + Source Sans 3)
data_gen/ # Synthetic 50-SKU x 4-system extract generator (seeded, deterministic)
dagster/ # Asset graph orchestrating generate -> load -> dbt -> export
dbt/ # staging -> intermediate -> marts; physics (cube, density, freight class) in macros
scripts/ # Standalone loaders and the frontend JSON exporter
config/ # cost_params.yml — every rate, volume, and tolerance, with source citations
frontend/ # React story app with client-side paradox recomputation
tests/ # pytest suites for cost math, data gen, export, and E2E reconciliation
MIT — see LICENSE.
Built by Lailara LLC — data hygiene and analytics consulting for specialty food brands scaling into national retail.