DATTS

The Future Doesn’t Prototype, It Generates: From Idea to Production in 48 Hours
A conceptual image representing the transition from human design to AI execution. The left side features traditional paper sketches and wireframes of a user interface on a rustic wooden table. The right side depicts a sleek, white humanoid robot with glowing blue accents working on a laptop, with lines of code visible on the screen.
Bye Bye Prototypes - Hello AI

How we built a full-scale, AI-powered industrial planning system — with a live dashboard, real optimisation engine, and 335 automated tests — in two days of conversation with Claude.

Let me tell you about the project that permanently changed how I think about software development.

A few weeks ago, I had an idea: could we build a real, production-grade planning system for a steel mill — the kind of system that normally takes a team of engineers six months to deliver — using only conversations with an AI? Not a toy. Not a demo. Not a mockup with fake data. A real working system.

Forty-eight hours later, we had one.

"The old way: spend weeks building a prototype to show the customer. The new way: build the real thing faster than the prototype would have taken."


This is the story of how we got there — and why it matters for every person in tech, business, and beyond.

The Old Way of Building Software

For decades, software development followed a predictable rhythm. A customer comes to you with an idea. You spend weeks — sometimes months — just figuring out what to build. Then you build a prototype. A static thing. Screens with fake data, no real logic, just enough to show the customer what it might feel like.

A project roadmap infographic detailing the stages and costs of software development. Stages include Discovery (2–4 weeks), Requirements (3–6 weeks), Prototype (4–8 weeks), Development (3–6 months), and Testing + Launch (4–8 weeks). A red summary at the bottom indicates a total project span of 6–12 months and a budget range of $200,000 to $500,000.
The Traditional Software Delivery Timeline

The prototype gets shown to the customer. They give feedback. The team goes back and redesigns. The customer waits again. By the time real code is being written, months have passed, and the market has moved.

And the worst part? The prototype itself is a lie. It's a painted wall — it looks like a door but leads nowhere. The customer can't touch real data with it. They can't run it. They have to imagine what the real thing would feel like.

Our Starting Point: A Half-Formed Idea

The project I had in mind was a Slitting Pattern Optimiser for steel manufacturing. Steel mills receive huge mother coils of steel — imagine a roll of paper towels, but each one weighs 20 tonnes and costs tens of thousands of dollars. These coils need to be slit (cut lengthwise) into narrower strips to fulfil customer orders. The cutting pattern matters enormously: waste even 3% of steel on every coil and you lose millions per year.

I came to Claude with an idea that was roughly this vague (this is what was in my mind)

"I want to build a steel mill slitting pattern optimizer using AI. It should help planners decide how to cut mother coils to fulfill customer orders with minimum scrap."


Thirty words. That's it. In the old world, this would kick off a multi-week "discovery phase" with workshops, user interviews, and requirements documents. Instead, what happened next took about four minutes. Claude generated a detailed (yet not fully elaborated, I must admit) prompt like this:

text
Function: Cutting / Slitting Pattern Optimization

Context:
Steel Hot Rolling Mills producing flat products.

Objective

Develop an optimal cutting and slitting plan to fulfill demand for narrow coils, sheets, or plates by efficiently utilizing available coil inventory while minimizing scrap and procurement requirements.

Inputs:
Mill Capabilities
* Narrowest possible rolling width
* Maximum rolling width
* Slitting and cutting limitations
* Permissible tolerances
* Configurable side trimming width
Inventory of Coils
* List of available parent coils with widths greater than the narrowest rolling width
* Relevant attributes such as thickness, grade, and quantity
Orders / Demand (Child Orders)
* Demand for narrow coils, sheets, or plates
* Required widths, quantities, and specifications
* Orders derived from wider parent coils
Processing Constraints
* Feasibility of slitting or cutting from larger coils
* Material compatibility and operational limitations

Outputs
Optimal Cutting / Slitting Plan
* Combine widths of child orders to generate efficient cutting patterns
* Maximize material utilization and minimize trim loss
Inventory Allocation
* Assign cutting plans to available coils to fulfill demand
Procurement Recommendation
* Identify unmet demand that cannot be satisfied using existing inventory
* Generate indents for wider coils with required specifications

Guardrails
* Utilize existing inventory as much as possible.
* Minimize scrap and side trim losses.
* Apply a configurable width for side trimming.
* Ensure feasibility within mill capabilities.
* Optimize for maximum yield and demand fulfillment.

Expected Deliverables
* A set of optimized cutting and slitting patterns
* Coil-to-order allocation details
* Scrap and yield analysis
* A list of unmet demands with procurement recommendations

How AI Turned 30 Words into a Product Specification

The prompt was good enough to start our journey!

An infographic titled "What AI Understood From 30 Words." The top section shows a workflow: Step 1 (30-word vague idea), Step 2 (AI domain mapping), and Step 3 (Full 12-section spec). The bottom section highlights six key technical insights inferred by the AI, including Coil Weight Constraints (12T–25T), CP-SAT Optimization using Google OR-Tools, and a Confidence-Gated Workflow for auto-approvals.
The Prompt Evolution Journey

The AI didn't just respond to what I asked — it anticipated what I hadn't asked. It knew that a steel mill optimiser would need grade substitution rules (a higher-grade steel can serve a lower-grade order, but not vice versa). It knew about remnant management (the leftover strips after slitting). It proposed confidence scoring so the system could route decisions to a human when it was uncertain. None of that was in my 30-word brief.

This is the first paradigm shift: AI doesn't just take instructions — it brings domain knowledge to the conversation.

Building It Together: Feature by Feature

I captured the context in detail in a Requirements Document. Once we had finalised the requirements document, we started building. Not sketching. Not wireframing. Actually building. Here is the complete journey, compressed into a single timeline:

A vertical list detailing an AI-driven development sprint over 48 hours. Milestones include building a FastAPI backend, integrating Google OR-Tools CP-SAT optimizer, establishing grade hierarchy rules for 28 steel grades, and generating SVG slit pattern diagrams. The timeline concludes with the creation of a KPI analytics engine, a live planner dashboard, and final production hardening with outlier data testing.
The 48-Hour Build Timeline

What We Actually Built

The numbers tell the story better than words.

Three cards displaying development metrics: 335 automated tests at 97% coverage, 14 fully documented API endpoint groups, and a 48-hour timeline from first prompt to live system.

But raw numbers don't capture the depth. Here is what the system actually does:

A grid of tags showcasing technical features of a manufacturing planning system, including CP-SAT optimization, KPI radar charts, steel grade management, and real-time dashboards.
A technical flowchart of a software architecture. It features four main blocks: a "Planner Browser" containing UI components like dashboards and modals; an "API Layer" powered by FastAPI and Python handling core services like inventory and analytics; an "Optimizer" utilizing Google OR-Tools CP-SAT; and a "Database" using SQLite with 12 relational tables. A red footer at the bottom confirms 335 passing tests and 97% coverage.
System Architecture - What was built

The Paradigm That Has Permanently Changed

The most important thing about this project isn't the steel mill. It's what it proves about how software gets built now. Let me put it directly side by side:

A comparison table titled "The Old Way" vs. "The New Way," showing a traditional 12-month software development cycle being replaced by a 48-hour AI-driven process from initial idea to production.
Paradigm Shift
From 6 months -> 48 hours.

But What About the Human?

The most common question I get when I share this story is: "So you just… talked to it?" And the answer is: yes, but that undersells the role of the human enormously.

Every decision that mattered was mine. When AI proposed a grade substitution algorithm, I reviewed the metallurgical logic and caught a nuance about stainless steel families that needed correction. When the confidence-scoring formula was proposed, I adjusted the weights. When the UI was built, I reviewed every screen and asked for changes. When bugs appeared — and they did — I understood what was wrong and directed the fix.

A text graphic explaining that AI-assisted development replaces "tedious mechanical work" while humans continue to provide essential domain knowledge, judgment, and final decisions.

The other thing that surprised me:AI makes you a better thinker.When you have a collaborator who can immediately implement whatever you propose, you start thinking harder about what to propose. I found myself asking questions I would never have bothered with before, because the cost of exploring them was effectively zero.

The Takeaway for Everyone in Technology

If you are a business leader, the budget you've been allocating for six-month software projects may need to be reconsidered. Not because the work is less valuable, but because it now costs a fraction as much to do it properly.

If you are a developer, the most valuable skill you can build right now is learning to direct AI effectively. Not to use it for simple tasks, but to collaborate with it on complex ones. The developers who will thrive are those who can hold an entire system's architecture in their head and guide AI through building it piece by piece.

If you are a domain expert in any field — manufacturing, logistics, finance, healthcare —you now have access to engineering capability that was previously only available to companies with large development teams. Your domain knowledge, combined with AI's implementation speed, is a competitive advantage that didn't exist two years ago.

"The prototype is dead. The real thing takes less time to build."


We built a system that a traditional team of four engineers would have taken six months to deliver. We built it in 48 hours of conversation. It has 335 tests, a live dashboard, a mathematical optimiser, a visual cutting diagram, and a complete API with documentation.

A screenshot of a web-based industrial dashboard titled "Slitting Pattern Optimizer." The interface features high-level KPIs including "150 Available Coils" and "30 Plans Awaiting Review." Middle sections detail configurations for three slitting lines (Heavy Gauge, Light Gauge, and Coated Products). The bottom contains data grids for Child Orders, Mother Coil Inventory, and a list of generated Slitting Plans with scrap percentage and cost metrics.
The Main Screen
A technical dashboard for slitting plan PLN-34D18001. The top section shows a visual coil layout with a blue 800mm strip and a green 650mm strip. Key metrics below include 97.3% utilization, 2.7% scrap, a confidence score of 89, and an effective width of 1490mm. A data table lists specific order details, and an "Approve Plan" button is highlighted in green at the bottom.
Screen displaying Slitting Pattern
An industrial analytics dashboard titled "KPI Performance & Radar." The left side shows performance cards: 4.1% Avg Planned Scrap, 100% Stock Fulfillment Rate, 67.3% Auto-Approval Rate, 95.9% Avg Utilization, and 0.0% Urgent Trigger Rate. On the right, a 5-axis green radar chart visually compares these actual metrics against a grey target zone for quick performance assessment.
KPI Dashboard
A dark-themed user interface for generating test data. The modal displays current database stats (230 coils, 610 orders) and offers five selectable scenarios. The "Normal" scenario is highlighted, showing a balanced dataset of 250 coils, 500 orders, and 40 remnants. A warning at the top notes that seeding will clear existing data unless specified otherwise.
The Test Data Generator for Automatic Testing

And we're just getting started.

My Honest & Humble Submission

Building the future of industrial planning, one conversation at a time. After all these years of building knowledge and experience, I’m convinced that what we once accepted as fixed development timelines is now being reimagined.

#ArtificialIntelligence #SteelIndustry #SoftwareDevelopment #DigitalTransformation #AI #Innovation #Manufacturing #Claude #ProductDevelopment #FutureOfWork