AI as a Velocity Multiplier for Domain Experts

I've Been a Professional Developer for 35 Years

Built systems through the dot-com boom, enterprise software, e-commerce platforms. Worked in marketing, operations, product—but I never stopped building. I'd see a problem and write code to solve it. Sometimes that was welcomed. Sometimes it wasn't.

The feedback was always the same: "Move fast and break things" works great until something breaks in production. Then people want proper error handling, comprehensive logging, defensive programming, documentation—all the discipline I knew I should have but would skip to ship faster.

When I bought this warehouse in 2024, I knew exactly what needed to be built. I'd been a client for nearly a decade. I'd watched the warehouse staff drown in paperwork. I'd seen clients struggle with batching complexity. I could build the solution—given 6 months.

But I didn't have 6 months. Clients needed it NOW.

---

The Dual Problem

This wasn't just a warehouse efficiency problem. It was two problems:

Problem 1: Client Batching Complexity

• Science Artifacts Client: Exports from custom order system to Excel, manually adjusts batch files, needs sequential numbering with collision detection for individually numbered specimens, warehouse filtering for split shipments
• Gaming Publisher Client: Generates CSV from ecommerce platform, tweaks in Google Sheets, needs intelligent auto-consolidation that groups exact SKU matches, auto-splits at 50 orders to prevent warehouse overwhelm, filters promotional items

Getting these complex rules into ShipStation wasn't simple. Each client had different workflows, different edge cases, different requirements. Off-the-shelf software couldn't handle it.

Problem 2: Warehouse Process Efficiency

• Staff spent 60% of time on paperwork (printing labels, sorting, matching)
• 100 orders = 100 individual labels, 100+ packing slips, separate paperwork per carrier
• Manual batch processing was error-prone
• Peak season required temporary hires just for paperwork

The solution was intelligent batch processing: consolidate 100 orders into 5-10 batches = 90% less paper. But it had to handle both clients' unique requirements AND make the warehouse more efficient.

---

The Old Tradeoff: Speed OR Professionalism

For 30 years, I'd lived with this tradeoff. When you're solo and need to ship, you choose:

Option A: Move Fast

• Skip comprehensive error handling
• No defensive null checks
• Minimal logging (debug when it breaks)
• Light documentation
• Ship in weeks

Option B: Professional Code

• Try/catch blocks everywhere
• Defensive if (s.tags && s.tags.length > 0) checks
• Comprehensive API logging
• Sequence gap detection
• Proper documentation
• Ship in months

I knew what professional code looked like. I'd written it in enterprise environments with teams and code reviews and QA. But solo? I'd cut corners to go fast.

---

The Continuous Development Approach

Here's what made this different from traditional development:

Week 1: 72 Hours to Production

Not perfect tools. Working tools. Basic batch processing, ShipStation API integration, initial workflows functional. Warehouse staff started using it immediately.

I could do this because:

• I knew exactly what needed to be built (deep domain knowledge)
• I knew I could fix what broke (30 years of debugging)
• Claude Code handled the professional discipline (error handling, logging, defensive checks)
• No external dependencies (no waiting on consultants or approval chains)

Weeks 2-6: Continuous Iteration

This is where the real power emerged:

Morning: Warehouse staff use the tools

Afternoon: They report what worked and what didn't

Evening: Improvements pushed to production

Not "planned for next sprint." Pushed that same day.

Deep Client Sessions:

• Watched clients work through their orders in real-time
• Saw where they got stuck or had to work around limitations
• Identified optimizations they didn't even know they needed
• Quickly written by Claude Code, tested same day

Live Warehouse Sessions:

• Watched staff use the batch files during actual fulfillment
• Saw visual/UX improvements that would save seconds per order
• Seconds × 1000 orders = real efficiency gains
• Refined same day

Everything focused on: speed, accuracy, efficiency.

---

What Actually Got Built

Let me show you the sophistication of what emerged from this approach. This isn't simple automation—it's intelligent logistics software.

Tool 1: Pre-Created Batch Processor (3,706 lines)

For the Science Artifacts Client

This client runs massive crowdfunding campaigns with individually numbered specimens. They export from their custom order management system to Excel, make manual adjustments (some items ship from HQ, not the warehouse), then import batch files.

What it does:

• 5-Step Guided Workflow: Import → Summary → Export → Labels → Warehouse files
• Sequential Numbering: Collision detection ensures specimen #1234 goes to the right backer
• Warehouse Filtering: Split shipments tagged "SHIP FROM HQ" automatically excluded
• Intelligent Summary: Groups batches by carrier/destination, shows ranges (e.g., "batches 1233-1240, 1245-1250")
• Process Automation: Creates batches in ShipStation, generates labels, exports warehouse-optimized PDFs

The client can modify their batch files in Google Sheets without any application overhead or WMS software limitations. The warehouse gets clean, sequential batch files optimized for their picking process.

Tool 2: Intelligent Auto-Consolidation (3,820 lines)

For the Gaming Publisher Client

This client ships board games and collectibles with complex SKU combinations and subscription boxes. They generate CSV from their ecommerce platform and want intelligent batching.

The Batching Intelligence:

Three Batch Types (Optimized for Warehouse Picking):

1. SINGLES: One unit per order (fastest picking)
2. MULTIS: Multiple units, all same SKU (efficient picking)
3. VARIOUS: Multiple different SKUs (requires attention)

Orders are automatically sorted in this sequence because warehouse staff can pick SINGLES and MULTIS faster, saving VARIOUS batches (which need more attention) for last.

Smart Consolidation Logic:

• Exact SKU Matching: Orders with identical items → consolidated into one batch
• Threshold Intelligence: If fewer than 3 orders match, dissolve and group into VARIOUS
• Auto-Splitting: Batches larger than 50 orders automatically split (prevents warehouse overwhelm)
• Promotional Filtering: Gift-with-purchase items excluded from batching logic

User-Configurable Settings:

• Minimum orders threshold for VARIOUS batches (typically 3)
• Include item details in VARIOUS batches (toggle)
• Order footer with semicolon-separated order numbers (toggle)
• Consolidation strategy on/off

This level of sophistication emerged from watching the client work through orders and the warehouse pick from batches. Each refinement was quick to implement because Claude Code handled the implementation details while I focused on the logic.

Shared Professional Features

Both tools include the discipline I'd normally skip when moving fast:

• Comprehensive API Logging: Every ShipStation call logged with timestamps, request/response data, automatic rotation to prevent sheet bloat
• Test Mode: Simulate entire workflow without making API calls (staff can practice risk-free)
• Error Recovery: Pause-and-retry for insurance funding errors, detailed error messages with recovery instructions
• Batch Renumbering Utility: Fix sequence gaps in production without downtime (processes in reverse order to avoid collisions)
• Status Tracking: Each batch sheet tracks creation status, batch ID, full batch name in fixed cells
• Defensive Programming: Null checks everywhere (ShipStation API is inconsistent)

---

The Business Results

Direct Measurable Impact

Paper & Printing: $850/month → $255/month = $7,140/year saved (70% reduction)

Staff Productivity: 2 FTE shipping 400 orders/day → 850 orders/day = 2.1x throughput (same headcount)

Error Reduction: ~8 wrong-warehouse shipments/month → ~0.5/month = ~$180/month saved in recall costs

Total Annual Direct Savings: $9,300

Strategic Advantages

• Peak Season Capacity: Can handle holiday rush without temporary hires
• Client Scalability: Clients can grow without worrying if fulfillment can keep up
• Faster Fulfillment: Friday orders ship same day instead of Monday
• Client Flexibility: New promotional bundle? Test and deploy same day (used to take weeks)
• Continuous Improvement: Can say "yes" to client needs instead of "let me check if our system can do that"

Development Cost vs. Alternatives

Traditional Custom Development: $15K-30K upfront, 3-6 months timeline, $500-1,000/month maintenance retainer

What I Did (Claude Code):

• Development: 40 hours × $50/hr = $2,000
• Maintenance: ~2 hours/month × $50 = $1,200/year
• Claude Code: $20/month = $240/year
• Total Year 1: $3,440

ROI: 270% ($9,300 savings / $3,440 cost)

Payback Period: 4.4 months

5-Year Savings: $46,500 (not including productivity gains or strategic advantages)

---

The Pattern (How You Can Apply This)

This approach works if you have these ingredients:

1. Deep Domain Knowledge

You need to know exactly what needs to be built. Not "I think this would help" but "I've done this process 1000 times and know every edge case." I'd been a client of this warehouse for a decade. I'd watched staff work. I knew the pain points intimately.

AI can't give you this. You bring this.

2. Development Experience

You need to know when AI gets stuck on dumb things. When it suggests over-engineered solutions. When defensive programming actually matters vs. when it's overkill. When to trust the code and when to debug.

This isn't "learning to code with AI." This is experienced developers using AI as a velocity multiplier.

3. Confidence to Ship Fast

I could go live in 72 hours because I knew I could fix what broke. That confidence comes from decades of debugging production systems. If you're not comfortable shipping imperfect code and iterating, this approach will stress you out.

4. Ownership of the Problem

The continuous development cycle only works when you own the problem end-to-end. I own the warehouse. I work with the clients. I watch the staff use the tools. I can iterate daily because there's no approval chain, no change orders, no external dependencies.

If you're building for someone else's business, this gets harder.

---

The Feedback Loop Changes Everything

Then: "Move fast and break things" got pushback because production breaks were expensive and professionalism took too long.

Now: You can move fast AND ship professional code. The discipline you know you should have (error handling, logging, defensive checks, documentation) doesn't slow you down anymore. AI handles it.

The feedback loop changes everything. Instead of:

• 6 months of development
• Launch
• Discover what clients actually need
• Schedule changes for next quarter

You get:

• 72 hours to working prototype
• Launch
• Discover what clients actually need
• Ship improvements same day

That tighter feedback loop means you build exactly what's needed instead of what you thought was needed 6 months ago.

---

Questions You Might Have

Q: Why Google Apps Script instead of a "real" backend?

A: Three reasons:

1. Client Accessibility: Clients modify batch files themselves in Google Sheets without application overhead or WMS limitations
2. Zero Ops: No infrastructure to manage, instant deployment, free hosting, built-in OAuth
3. Right Scale: 500-1000 orders/day doesn't need microservices. It needs something that works and stays out of the way.

Q: How much is Claude Code vs. your code?

A: ~80% Claude Code generated, 20% me debugging and adding domain knowledge. But I wrote every prompt, reviewed every line, and knew exactly what to ask for.

Q: Would this work for someone learning to code?

A: Probably not. You need to know when AI is wrong, when to trust it, and when to override it. That judgment comes from experience.

Q: What about Claude Code version?

A: Started on Sonnet 3.5, now using Sonnet 4.5. The velocity multiplier keeps getting better.

Q: Can I see the code?

A: The repository is private (contains client data), but the full case study has architecture details and technical specifications.

---

The Bottom Line

This isn't "AI will replace developers." It's not "anyone can code now."

It's: Experienced developers with deep domain knowledge can now build sophisticated systems at 5x velocity without sacrificing professionalism.

The constraint used to be: "I know what to build, but it'll take 6 months."

Now it's: "I know what to build. I'll ship a working version this week and iterate based on real use."

That changes everything.

---

Built in 3 weeks (November 2024) in Upper Tract, West Virginia. Production since November 2024 with 99.5% uptime.

---