LifeSeasons Warehouse Batching Optimization

Part 1 — Section 02

Cart Types & Capacities

LifeSeasons uses seven cart types. All capacities, bin counts, and physical cart quantities are now confirmed. Dynamic and Distributor are identified but sit outside the main batching algorithm.

Green Cart (5×9)

45 bins

Small multi-product orders through the auto-bagger. 9 carts in warehouse. (confirmed)

Per bin: 3 bottles

Details ▾

3 bottles per bin due to current bag size — confirmed. Cannot include restricted SKUs (272, 284, 503, 307, 317, 369) — these are too large for the auto-bagger. Classification proxy: ≤3 unique shelf locations AND ≤3 total bottles. SKU 505 is NOT on the exclusion list (confirmed by Sherry).

Red Cart (2×5)

10 bins

Large orders, 29–80 bottles. 10 carts in warehouse. (confirmed)

Per bin: max 80 bottles

Details ▾

Sherry stated explicitly: “80 bottles is our limit to how many bottles can fit in one red bin.” Orders with >80 bottles fall back to the mailer cart. 10 physical carts in the warehouse.

Mailer Cart

80 bins regular

NQ-prefix & mailer SKUs, or bulk fallback (>80 bottles). 12 carts in warehouse. (confirmed)

Per bin: infinite

Details ▾

Regular mailer: 80 bins, 12 carts. No physical bottle limit per bin. Triggers: orders containing the mailer insert SKUs (9085, 9136, 9137, 9138). NQ-prefix orders only go on mailer if the mailer + insert SKU is in the order — otherwise they route by size (bag or basic box). Also the overflow fallback for any order >80 bottles. SKU 505 routes to mailer (yellow cart removed).

NQ Pre-Kit Cart

120 bins

Pre-kitted NeuroQ orders. 12 carts in warehouse. (confirmed)

Per bin: infinite

Details ▾

Larger pre-kit variant of the mailer cart: 120 bins per cart, 12 carts. No physical bottle limit per bin. Used for pre-kitted NeuroQ orders specifically.

Pack-One Cart

200 bins

Single-product orders, 1 line, 1–3 bottles. 6 carts in warehouse. (confirmed)

Per bin: 1–3 bottles

Details ▾

Fastest pick type — every bottle is the same product so the picker grabs a pre-prepped box. 1 line item, 1–3 bottles per bin, 200 bins per cart, 6 carts. Our proxy: exactly 1 unique shelf location AND ≤3 total bottles.

Blue Cart (4×7)

28 bins

Standard multi-product orders up to 28 bottles. 6 carts in warehouse. (confirmed)

Per bin: ≤28 by rule

Details ▾

Standard: Orders with ≤28 bottles. Max 28 per bin by rule. 6 physical carts in warehouse.

Vertical config: SKUs 272 and 284 (large boxes, can’t go in auto-bagger) use all 7 bins in a single column for one order.

Dynamic Cart

1 bin

Exception / hold cart for orders outside the standard batching flow. 4 carts in warehouse. (confirmed)

Per bin: infinite

Details ▾

1 bin per cart, no bottle limit per bin, 4 carts in warehouse. Used as a manual hold — orders are assigned here when they need extra attention, isolation, or a holding pattern. This is always a human decision; it cannot be automated. Not included in the batching algorithm.

Distributor Cart

100 bins

Wholesale distributor orders. Detected via CustomField2. 3 carts in warehouse. (confirmed)

Per bin: infinite

Details ▾

100 bins per cart, no bottle limit per bin, 3 carts in warehouse. Identified by CustomField2 = “Distributor Order” or “Distributor Health Concerns” in the WMS export. Classified first in the hierarchy (step 1). Wholesale only — not DTC. Assigned to specific people by Sherry.

Auto-Bagger Exclusions — Confirmed List

These SKUs cannot go through the auto-bagger (too large or boxed). They must be routed to blue or red carts instead of green. Confirmed by Ben and Sherry (SKU 505 is not on this list — confirmed by Sherry):

SKU 272 SKU 284 SKU 503 SKU 307 SKU 317 SKU 369

Part 1 — Section 03

Sherry's Current Assignment Process

Sherry runs through these filters every morning in order. Each step pulls a group of orders and assigns them to a cart type. Whatever remains after all filters goes through size-based rules at the end.

Assignment happens in this exact sequence — order matters

Each step removes orders from the pool. Later steps see only what remains.

1

SLA / Priority Check

DTC orders (Shopify store in CustomField4) that came in before 3pm must ship today — processed first, every morning. After 3pm DTC goes next business day. Wholesale always goes next business day.

2

Brand / Channel Routing

NeuroQ (NQ order prefix, SKUs 9136/7/8) → mailer cart. Symmetry (SKU 9085) → mailer cart. My Best Heart (SKU 505/503) → yellow if high volume, else blue. NGVC → grouped together (Speed Cell 3). HC → grouped together, never with NGVC. Distributor orders → manually assigned separately.

3

Remove Restricted SKUs

SKUs 272, 284, 503, 307, 317, 369 cannot go in the auto-bagger. Pulled out before green cart assignment. SKUs 272 and 284 (large boxes) go on blue carts in vertical 4×7 configuration.

4

Pack-One Check

Every bottle in the order is the same product → pack-one cart (1–12 bottles). Fastest pick type — the worker grabs a full pre-prepped box from the pallet rather than hunting individual shelves.

5

Size-Based Assignment (remaining orders)

2–3 items, fits small bag → Green cart. Up to 28 bottles → Blue cart. 29–80 bottles → Red cart. >80 bottles → Mailer cart (bulk fallback). Blue can go vertical (4 orders × 7 bins) for larger individual orders. FIFO ordering on blue — oldest orders ship first.

Sherry also tries to create full carts wherever possible — a partial cart is inefficient because a worker walks the same distance for fewer orders. At end of day, leftover orders may go on any available cart rather than waiting.

Part 1 — Section 04

Brand & Priority Breakdown

From the 930 usable orders in the ~10-day sample. Brand is detected from item number prefixes. Orders from different brands are never grouped into the same batch.

Pack-One Cart

483 orders (52%)

Proxy: 1 unique shelf location, qty ≤12. Capacity 44 (derived from historical data)

Green Cart

329 orders (35%)

2–3 distinct products, no restricted SKUs. Capacity 45 (5×9 cart) (confirmed from WMS)

Blue Cart

102 orders (11%)

Up to 28 bottles. Capacity 28 (4×7 cart) (confirmed)

Red Cart

13 orders (1.4%)

29–80 bottles. Capacity 10 bins (2×5 cart) (confirmed)

Mailer Cart

3 orders (<1%)

Rule: NQ-prefix orders or SKU 9085 (Symmetry). These 3 are >80-bottle fallback orders — NQ and 9085 orders not in sample. (partial)

Yellow Cart

0 orders

Rule: SKU 505 (My Best Heart, crushable). Batched on yellow when volume is sufficient, otherwise falls to blue. Capacity 40 (5×8 cart) (confirmed from WMS). SKU 505 not in sample — untestable.

Brand Breakdown

Detected from item number prefixes. Orders from different brands are never grouped into the same batch.

Health Concerns

44 orders

Items start with HC. In our sample all are wholesale (UPS Ground). Ben listed HC as a D2C channel — needs clarification (see Q3 below).

NGVC / Private Label

54 orders

Items start with PL-NGVC. Natural Grocers Vitamin Cottage. Stored on Speed Cell 3. Never mixed with HC.

LifeSeasons / Other

832 orders

All other brands (Femitry, Inner Eco, standard LS). Can share batches with HC or NGVC per Ben, but HC and NGVC never share a batch.

Priority Breakdown

Wave 1 — Urgent (DTC)

570 orders

Hyphenated order numbers (e.g. 4046858-3990825). All carry OrderLogix in CustomField4 in current data. Need updated export with Shopify store names.

Wave 2 — Standard

360 orders

Plain numeric order numbers. Wholesale (Sage) or blank CustomField4. UPS Ground shipping.

Part 1 — Section 05

Open Questions for LifeSeasons

Four gaps need clarification before the system can go live. Everything else has been confirmed.

1

What is the actual bottle threshold between blue and red?
Our current rule: orders with ≤28 total bottles go on blue; 29–80 bottles go on red. But in the WMS data, 184 orders that WMS put on a red cart have ≤28 bottles — our system would route those to blue. This means either our 28-bottle cutoff is too high, or there’s a secondary condition we don’t know about that pushes lower-bottle orders to red.

What is the exact maximum bottle count before an order moves from blue to red?
2
Pack-one: is the rule SKU-based or location-based — and does it overlap with green?
Our proxy rule: exactly 1 shelf location AND ≤3 total bottles = pack-one. But in the WMS data, 62 orders we call pack-one went on blue and 37 we call pack-one went on green. Two things need clarification:
- A: Does pack-one require every item in the order to be the same SKU/product — or is “1 shelf location, ≤3 bottles” enough regardless of how many SKUs?
- B: If an order has 1 shelf location and only 1–3 bottles, should it go to pack-one or green? Our hierarchy sends it to pack-one, but WMS data suggests some of those go green.
3

If an order has both a blue-forced SKU and a mailer SKU, which cart does it go on?
SKUs 272 and 284 (large boxes) force an order to the blue cart. SKUs 9085, 9136, 9137, 9138 force an order to the mailer cart. If a single order contains both types — for example, a large box (272) plus a Symmetry bottle (9085) — our system currently sends it to mailer because the mailer check runs first in the hierarchy.

Is that correct, or should the large-box rule override mailer?
4

What determines whether a mailer run uses the regular cart (80 bins) or the pre-kit cart (120 bins)?
We have confirmed capacities for both: regular mailer = 80 bins, pre-kit mailer = 120 bins. But the algorithm currently always uses 80 as the limit because we don’t know which orders are pre-kitted.

Is pre-kitted determined by specific SKUs or the order channel — or is it a manual decision made by the team each run? If it’s manual, the system just needs a way for ops to indicate which capacity to use when releasing a batch.

Part 1 — Section 06

Cart Assignment — How the System Runs

This is the full pipeline from raw WMS export files to each order being assigned a cart type. Every decision is made in the sequence shown below — top to bottom, left to right.

📥

Input — WMS Export Files

Two text files exported from the LifeSeasons WMS every morning.

Inventories file

Every shelf location in the warehouse + its pick sequence number (1,563 locations in sample)

History Logs file

Every pick event — order number, item number, shelf location, quantity, batch ID

📄

Parse & Load — real_data_parser.py

Raw text files are read line by line and turned into usable data structures.

load_inventory()

Outputs a location map: shelf name → pick sequence number. Used to order the warehouse route.

load_history()

Filters to PickComplete rows. Outputs a table of order rows with order number, item, location, qty, priority.

📦

Build Orders — build_orders()

Groups all pick rows by order number. For each order, computes: total bottle quantity, unique shelf locations, brand (HC / NGVC / LS), priority (urgent D2C or standard wholesale), and which SKUs it contains.

930 orders extracted 570 urgent (DTC) 360 standard (wholesale)

🛒

Cart Classifier — 8-Step Hierarchy

Each order passes through these checks top to bottom. The first match wins. Remaining orders fall to the next step.

1

Mailer cart — NQ order prefix or SKU 9085 (Symmetry). Cap 80. mailer

2

Yellow cart — SKU 505 (My Best Heart, crushable). Cap 40 (5×8). Falls to blue if volume low. yellow

3

Blue-forced — SKUs 272 or 284 (large boxes). Must go on vertical blue cart (4×7). Cannot go in auto-bagger or any smaller cart. blue (forced)

4

Pack-one cart — Real rule: every bottle is the same product, 1–12 bottles. Proxy we use: exactly 1 shelf location AND ≤12 bottles (we can't see product identity in the data). Cap ~44 (derived from data, unconfirmed). pack_one

5

Green cart — Real rule: 2–3 distinct products, fits small bag. Proxy we use: ≤3 unique shelf locations AND no auto-bagger-excluded SKUs (272, 284, 503, 307, 317, 369, 505). Uses auto-bagger. Cap 45 (5×9). green

6

Blue cart — Total bottles ≤28. Standard upright bins. Cap 28 (4×7). blue

7

Red cart — Total bottles 29–80. Large bins, up to 80 bottles per bin. Cap 10 bins (2×5). red

8

Mailer fallback — Total bottles >80. Bulk orders too large for any cart. mailer

⥇

Yellow Fallback Check — apply_yellow_fallback()

After classification, count all yellow-cart orders. If fewer than 5 MBH orders exist today, reclassify them to blue — not enough volume to pull a dedicated yellow cart. Threshold (5) is unconfirmed; need to ask Sherry.

🔄

Brand & Priority Split

Each order now has: cart_type + brand (HC / NGVC / LS) + priority (1=urgent, 2=standard). These three together define which batching pool the order belongs to. HC and NGVC orders are never placed in the same batch. Priority 1 (DTC) is processed before priority 2 (wholesale).

HC

44

orders

NGVC

54

orders

LS / Other

832

orders

✓

Output — Orders Assigned to Cart Types

Every order now has a confirmed cart type, brand label, and priority. This is the input to the batch optimization algorithms in Part 2.

483 pack_one

329 green

102 blue

13 red

3 mailer

0 yellow

Part 2 — Section 01

What Is Batch Optimization?

Once orders are assigned to a cart type (Part 1), there is a second decision: within that cart type, which specific orders get grouped into one batch? For example, if you have 100 blue cart orders, you need roughly 4 batches of 28. Which 28 go together in each batch?

Sherry does not currently optimize this. She assigns orders to batches based on what fits — not on which orders are physically closest on the shelves. This is the gap our algorithms fill. Grouping orders that pick from nearby shelf locations means the worker walks a shorter total distance per batch.

930

Orders tested

~10 days of real LifeSeasons data

1,563

Shelf locations

Mapped with pick sequence numbers

4

Algorithms tested

vs. WMS and random baseline

Part 2 — Section 02

Route Cost Model

We do not have the physical warehouse floor plan yet. As a proxy, route cost is:

Route Cost = 2 × max pick sequence in batch Walk from depot (seq = 0) to the farthest shelf location and back. Applied identically to all methods including WMS.

Known limitation: Pick sequence numbers do not capture cross-aisle travel or real warehouse geometry. The relative ranking of algorithms is reliable. The absolute improvement percentages will change once we have real walking distances from the floor plan.

Part 2 — Section 03

The 4 Algorithms We Tested

Each algorithm runs independently on orders of the same cart type, brand, and priority. Green cart orders are never mixed with blue cart orders. HC orders are never mixed with NGVC orders.

Baseline

Jaccard Greedy

Seed batch with farthest-location order
Fill by shared location overlap score
Fast, simple, good comparison baseline

Enhanced

RAG — Routing-Aware Greedy

Same seed as Jaccard
Screen top-15 by Jaccard similarity
Pick the one that adds the least route cost

Winner

Clarke-Wright Savings (CW)

Calculate savings from merging every pair of orders
Merge highest-savings pairs first
Industry-standard vehicle routing algorithm
Best performance across all metrics

Variant

CW + Consolidation

Run CW, then merge small batches to fill carts
Slightly worse than CW alone with smaller cart sizes
Not recommended for current capacity profile

Part 2 — Section 04

Results

All numbers use real LifeSeasons data (930 orders, ~10 days). Per-cart-type capacities applied. Brand and priority grouping enforced. Lower average route cost = less walking = better.

Average Route Cost Comparison

Lower bar = less walking = better. Scaled to Random baseline (59,118).

Random (no system)59,118

Baseline

no optimization — random grouping within cart types

WMS (current system)51,901

+12.2% vs Random

current production — modest improvement over random grouping

Jaccard Greedy49,944

+15.5% vs Random | +3.8% vs WMS

location-overlap grouping

RAG (Routing-Aware Greedy)49,944

+15.5% vs Random | +3.8% vs WMS

marginal cost minimization

Clarke-Wright Savings Best43,439

+26.5% vs Random | +16.3% vs WMS

pairwise savings merging — clear winner

CW + Consolidation46,467

+21.4% vs Random | +10.5% vs WMS

slightly worse than CW alone — consolidation merges orders from different shelf areas

Method	Avg Route Cost	vs Random	vs WMS	Status
Random (no system)	59,118	baseline	—	No optimization
WMS (current)	51,901	+12.2% better	baseline	In production
Jaccard Greedy	49,944	+15.5% better	+3.8% better	Tested
RAG (Routing-Aware Greedy)	49,944	+15.5% better	+3.8% better	Tested
Clarke-Wright Savings	43,439	+26.5% better	+16.3% better	Top Performer
CW + Consolidation	46,467	+21.4% better	+10.5% better	Tested

Clarke-Wright is the clear winner — 16.3% better than the current WMS on average route cost. Note that the WMS was free to mix priority 1 and priority 2 orders in the same batch, while our algorithms are constrained to never mix priorities. Despite this handicap, CW still wins by 16.3%, making the result stronger than it appears.

Part 2 — Section 05

Limitations & What's Next

What We Need for Part 2 to Be Exact

1
Warehouse Dimensions
- How long is one aisle (shelf row)?
- How wide is the gap between two adjacent aisles?
- Where does a picker start and end their route?
2
Warehouse Floor Plan or Map
- Even a rough hand-drawn layout would let us calculate real aisle-level distances.
- This is the single most impactful input for making optimization numbers exact and giving LifeSeasons a confirmed dollar figure on labor savings per day.

Current State vs After Answering Open Questions

Right Now

— Pick-sequence proxy for distance

— ~10 days of historical data

— Pack-one / green / mailer capacity estimated

— Cart assignment rules partially confirmed

✓ CW already 16.3% better than WMS

After Open Questions Answered

✓ Confirmed cart classification for all order types

✓ Exact capacities for every cart type

✓ Accurate D2C / SLA detection with order creation time

✓ Real warehouse geometry for exact distances

✓ Confirmed dollar figure on labor savings per day

Cart Types & Capacities

Auto-Bagger Exclusions — Confirmed List

Sherry's Current Assignment Process

Brand & Priority Breakdown

Brand Breakdown

Priority Breakdown

Open Questions for LifeSeasons

Cart Assignment — How the System Runs

What Is Batch Optimization?

Route Cost Model

The 4 Algorithms We Tested

Jaccard Greedy

RAG — Routing-Aware Greedy

Clarke-Wright Savings (CW)

CW + Consolidation

Results

Limitations & What's Next

The Distance Problem

What We Need for Part 2 to Be Exact