No Spec, No Problem: How FLS Turned a Two-Page Brief Into a Production-Ready AI System
“The result is faster, with more consistent outcomes and reduced costs,” said founder Collin Sedmak. “We can capture and scale critical expertise, turning decision-making into a repeatable system that can be applied across domains.”
Challenge
Heuritex came to First Line Software with a high-level idea: apply their existing mathematical methodology and a layer of AI to automate purchase order prioritization for inventory planners.
The problem was clear. Each week, supply chain personnel would open their ERP system, review a list of open POs, and manually determine the right action for each one—with no decision support, no ranking, and no way to capture or learn from past decisions. The process was slow, inconsistent, and entirely dependent on individual expertise.
What Heuritex wanted was a system that could do three things:
Apply their own scoring rules (owner-configured weights across item preference, margin, supplier trust, and historical pattern)
Layer in AI recommendations that improve over time as outcomes are tracked
Present all of this as a decision-support tool that empowers the planner — not one that replaces them
The challenge for FLS was that none of this existed yet, even as a specification. The entire brief was a two-page PDF describing the methodology.
Approach
With only a methodology document, FLS applied RACE Mode — an AI-native development approach designed to turn intent into a production-ready system.
Phase 1 — Strategic Scoping via AI Dialogue
Rather than waiting for a formal requirements process, the ACE Engineer uploaded the client’s PDF to Claude Code and spent several hours interrogating it — asking what the information was sufficient to build, what a meaningful proof of concept would need to demonstrate, and how to make the AI learning curve visible and credible to a skeptical audience. This AI-assisted scoping session replaced what would traditionally require a business analyst, a discovery workshop, and multiple client calls.
Phase 2 — PoC Development (Under 10 Hours)
The ACE Engineer built a fully clickable prototype in approximately 5–6 hours. Because no real client data was available, synthetic data was generated to simulate the system’s behavior over time — including a “time machine” view that stepped through simulated days and weeks to show the AI learning curve improving as more decisions were tracked. The goal wasn’t to show a finished product; it was to prove the path.
Phase 3 — Client Presentation Before Pricing
The prototype was demonstrated to the client before any pricing or timeline discussion. The client, who had shared their idea only days earlier, saw a working visual demonstration of how their concept could operate in practice. Their reaction confirmed the product direction and unlocked the information needed to build the real application.
Phase 4 — Application Build
After the client call, the ACE Engineer built a near-production-ready application over several days — incorporating the additional details gathered during the demo. Claude Code generated not only the application code but also user-facing documentation and a methodology guide, which the engineer used to validate their own understanding of the implementation.
What made this approach work:
- Starting from intent, not specification — the PDF was enough
- Using AI to accelerate scoping, not just coding
- Showing before telling — the prototype built trust before any commercial conversation
- Treating the engineer as product owner, making UI/UX decisions rather than waiting for requirements
- Compressing the classic idea → prototype → validation → build cycle from months to days
Solution
FLS delivered a web-based PO prioritization and decision-support application that operationalizes Heuritex’s proprietary scoring methodology and introduces an adaptive AI layer on top of it.
Core capabilities include:
Three-voice decision support: Each purchase order is evaluated through three lenses: the owner’s heuristic recommendation (based on configured preference weights), an AI-generated recommendation that learns from tracked outcomes, and the planner’s final decision — all displayed side by side.
Heuristic scoring engine: The system applies Heuritex’s scoring model using owner-configured weights, producing a linear composite score for each open PO. This score reflects item criticality, supplier reliability, cost/capital pressure, and historical risk patterns.
Adaptive AI layer: A lightweight AI model blends historical outcome data with the linear heuristic scores. As more planner decisions are tracked and outcomes are logged, the AI’s influence on final rankings increases. In the early stages, the system is primarily heuristic-driven; over time, it becomes increasingly informed by what has actually worked.
Governance overlay: Deterministic rules sit on top of the scored output — flagging critical items with near-term backorder risk for immediate review, and surfacing capital exposure information for lower-priority POs.
Planner-facing output: Planners see a prioritized list with plain-language action recommendations (expedite, monitor, clear, create new PO) and the reasoning behind each. The system captures each decision for future AI training and allows decision exports for manual ERP action in the pilot phase.
Planned future capabilities: Direct ERP integration for real-time data ingestion; automated action execution based on planner decisions.
Technology
The stack was selected to support rapid delivery, maintainability, and clean separation between the scoring engine and user-facing application, all while keeping infrastructure lean for an early-stage product.
AI-assisted delivery
- Claude Code: used for scoping, application code generation, documentation, and mockup creation
Core stack
- React 18 with TypeScript (frontend)
- Python / FastAPI (backend API)
- PostgreSQL (production database)
Reliability and visibility
- Vitest, Playwright, and pytest (frontend, end-to-end, and backend test coverage)
- GitHub Actions (CI/CD — automated lint, test, build, and deploy pipeline)
Infrastructure
- Docker Compose (containerized services)
- Hetzner VPS on Ubuntu (production hosting)
- Caddy (reverse proxy and HTTPS)
Results / AI Impact
Note: The application was approximately 90–95% complete at time of writing and had not yet been deployed with real client data. Measurable business outcomes are not yet available. The following reflects validated delivery outcomes.
Speed of delivery
A production-ready application was built by a single ACE Engineer, starting from a two-page PDF, with no prior domain knowledge of inventory management or machine learning. The initial PoC was completed in under 10 hours. A near-production application followed within days of the first client call.
From manual process to structured decision support
Before this system, inventory teams reviewed open POs manually each week with no recommendation engine — relying entirely on individual expertise and ERP data with no prioritization logic. The new system replaces that process with a scored, ranked, AI-assisted workflow that captures every decision for continuous improvement.
Proposal delivered from a working system
Rather than building a prototype to sell and then rebuilding for production, RACE Mode produced a near-finished application before the commercial proposal was written. The mockup-style proposal was generated from real screens — meaning the “estimate” reflected a system that already largely existed.
A platform ready to learn
Because every planner decision is tracked, the adaptive AI layer has a mechanism to improve from the first day of real use. The system is designed to get more useful over time, not just function at a fixed capability level.
Final Takeaway
Heuritex didn’t come to FLS with a specification. They came with an idea and a methodology. RACE Mode turned that intent into a working, production-ready system — before any formal requirements were written, before pricing was discussed, and before most teams would have finished their first discovery call.
This is what AI-native delivery looks like when the bottleneck isn’t coding. It’s defining intent clearly enough for a human-AI team to build the right thing with urgency.
Is This Approach Right for You?
RACE Mode works best when:
- You have a clear business problem and a methodology — but no formal specification
- You need to validate product direction before committing to a full build
- Speed to proof matters more than process compliance
- You’re an early-stage company, a corporate innovation team, or a founder who needs to show something real to stakeholders, investors, or clients — fast
If you can articulate the intent, FLS can build the system.