An AI Quoting Revolution Born on the Factory Floor
For non-standard CNC machining, quoting has long relied on veteran masters’ intuition—from part interpretation and process planning to cost calculation—all manual, taking 2–3 hours per quote. In 2026, Cao Dongdong, a factory professional with 15 years of experience, completed the 12th iteration of his AI quoting system, Union, on the AMD Ryzen AI Max+ 395 platform, winning the Professional Track OPC Solo Company Champion at the AMD Ryzen AI Agent Innovation Contest.
- Hardware platform: AMD Ryzen AI Max+ 395 (16-core/32-thread Zen 5 CPU + Radeon 8060S GPU + 50 TOPS NPU + 128GB LPDDR5x unified memory)
- Core model: ~27-billion-parameter dense model (neither a 7B small model nor a 70B+ large model)
- Architecture: Local inference, supporting offline deployment
- Current status: Deployed in seed testing at 3 factories
Multi-Agent Assembly Line: Decomposition Without>Invention
Cao stresses the system does not let LLMs “guess” prices but builds an interpretable, auditable digital pipeline. The system is decomposed into multiple agents: drawing parsing, DFM (Design for Manufacturing) analysis, process evaluation, quality inspection, finance, and risk control.
Key design counter-intuition: Despite multi-agent architecture, to avoid resource contention, the system adopts hybrid execution—some tasks run in parallel, yet results return sequentially—balancing throughput and stability. Meanwhile, agents exchange structured fields, not raw text, drastically cutting context length and token consumption.
The workflow follows strict layering:
- Parsing layer: STEP file parsing and feature extraction handled by CPU
- Decision layer: Multi-path process and tool selection由 local LLM
- Deterministic layer: Material usage, labor hours, and cost summation_by Python formulas
Engineering Tradeoffs Behind Hardware Selection
AMD Ryzen AI Max+ 395 was chosen for four concrete reasons:
| Selection Criterion | Rationale |
|---|---|
| Unified memory capacity | 128GB unified + up to 96GB configurable VRAM; 4-bit quantization fits a ~27B model entirely |
| ROCm ecosystem | Supports PyTorch and llama.cpp, lowering local deployment barrier |
| GPU performance | Integrated GPU delivers 256GB/s memory bandwidth, reducing inference latency |
| Industrial compatibility | Native STEP and CAD support, enabling integration into existing workflows |
Cao deliberately avoided larger models. The sweet spot lay in balancing memory footprint, generation speed (~20–30 token/s), context length, and tool-calling stability. Though too slow for high-concurrency online services, quoting time dropped from hours to ~3 minutes for simple parts versus 2–3 hours manually.
Real-World Challenges from Demo to Product
Deployment hurdles extend well beyond model adaptation. After migrating from Linux to Windows,现场 revealed multiple pain points:
- Some clients run legacy Windows Enterprise with aggressive security suites blocking containers and drivers
- STEP files suffer inconsistent encoding; robust fallback parsing and field-forecasting are essential
- Local deployment shifts complexity onto developers: what cloud APIs hide in black boxes now surface as driver and dependency onus
Factory-specific quoting habits differ sharply: one specializes in stainless steel, another in aluminum; equipment, tool inventory, labor costs all vary. Ten initial templates cover ~80% of base scenarios; final calibration requires historical order data and live feedback.
Deployment Recommendations
- Adopt now if: You run a non-standard CNC shop, have moderate order volume, require on-premise图纸 processing, operate in unstable or offline-capable environments
- Wait if: You need strict low-latency or high-concurrency quoting SaaS, or your parts are highly standardized with minimal process reasoning needs
The Bottom Line
LLMs haven’t replaced师傅—they’ve transformed their tacit knowledge into inspectable, iteratable processes. The Ryzen AI Max+ 395 provides a local compute substrate housing models, knowledge bases, and legacy industrial tools, while OPC super-individuals leverage this platform to redefine how factory intelligence is packaged and delivered.