August 15, 2026
Pacific Glazing Corporation | August 15, 2026 | Prepared for: Steve Watts, CEO
The AI deployment landscape has shifted decisively: the unit of value is no longer the model itself but the skill—packaged domain expertise that can be installed, shared, and executed. This creates a direct opportunity for PGC to convert its proprietary glazing knowledge into a defensible asset before the ecosystem consolidates. Edge inference has reached production viability, making field deployment technically feasible. The strategic risk is not that the technology is unavailable, but that competitors may lock up domain-specific skill libraries first.
What it is: AI capabilities are now distributed as installable "skills" or "plugins" rather than standalone models. Platforms like DeepSeek Harness (100K+ GitHub stars) function as app stores for domain-specific capabilities.
Why it matters: Whoever builds and controls skill libraries controls the deployment layer. PGC can build glazing-specific skills that competitors cannot easily replicate—similar to how specialized mobile apps created moats in the 2010s.
What it is: Running AI models locally on field devices rather than relying on cloud connectivity. Tools like h3.c (C-based inference optimized for specific hardware) demonstrate that local AI works on current hardware.
Why it matters: Field technicians often work in areas with poor connectivity. On-premise AI enables real-time decision support without latency or connectivity dependencies.
What it is: General-purpose AI models hallucinate and fail on specialized tasks. Grounding models on proprietary data—defect imagery, measurement protocols, installation workflows—is necessary for reliable field deployment.
Why it matters: Generic AI will not accurately assess glass defects or guide complex installations. PGC's historical data, if properly formatted, becomes the training substrate for reliable AI tools.
What it is: Emerging tools like Vero (verified code generation) and QuoteBench (command-path failure analysis) treat AI outputs as requiring mathematical proof rather than just testing.
Why it matters: Regulated industries will demand verifiable AI correctness. For glazing, this means liability-proof documentation of AI-assisted decisions—reducing risk if an installation fails.
What it is: DeepSeek Harness's plugin ecosystem is predominantly Chinese-developed. Western analysts systematically undercount this parallel AI platform economy.
Why it matters: PGC may find valuable glazing automation skills emerging from Chinese developers. Ignoring this ecosystem creates blind spots in competitive positioning and partnership opportunities.
What it is: The shift from "run models locally" to "optimize models for your specific hardware" is underway. Apple Silicon's unified memory architecture and tools like h3.c enable hardware-tailored inference.
Why it matters: PGC's field devices (tablets, measurement tools, mobile workstations) have specific hardware profiles. Optimization for those profiles determines whether edge AI is fast enough for real-time use.
What it is: Emerging research enables AI systems to reason about where and when operations occur—location-aware and time-aware decision-making.
Why it matters: Glazing installation is inherently spatial and sequential. AI that understands building geometry, local building codes, and installation sequence constraints will significantly outperform generic assistants.
What it is: Current proliferation of skills across Claude Code, DeepSeek Harness, and custom frameworks mirrors JavaScript framework chaos in the 2010s. No cross-harness abstraction layer exists.
Why it matters: Building skills for today's dominant platform may create lock-in. PGC must architect for portability—treating skill formats as temporary, not permanent standards.
What it is: The book-to-skill pattern enables conversion of static documentation (manuals, standards, training materials) into executable AI skills.
Why it matters: PGC has decades of written installation procedures, safety standards, and technical documentation. This content can become the foundation of a glazing skill library without starting from scratch.
What it is: Research on DRAM scrambling demonstrates hardware-level exploitation that bypasses CPU security boundaries, potentially undermining cloud-based confidential computing.
Why it matters: If verified, this undermines trust in cloud AI for sensitive workflows. Edge deployment becomes not just a performance choice but a security requirement for protecting proprietary processes.
Skills + Proprietary Data + Edge Inference = Field Deployment Ready The convergence of these three trends creates the technical foundation for PGC to deploy AI to field technicians. Skills provide the executable format, proprietary data ensures reliability, and edge inference enables offline operation.
Skills + Ecosystem Fragmentation = Architecture Risk The value of building skills is clear, but the lack of standardization means PGC must invest carefully. Skills built today may require migration tomorrow. Architecture decisions matter more than immediate tool selection.
Spatiotemporal Composability + Domain Grounding = Operational AI Generic spatial reasoning is insufficient for glazing. When spatiotemporal capabilities are grounded in PGC's specific measurement data and installation sequences, the result is operational AI that understands the physical constraints of the job.
Verification + Cloud Security Risks = Edge Preference Formal verification requirements push toward on-premise AI, and cloud security research provides another reason to prefer edge deployment. For PGC, this reinforces the strategic direction toward local inference rather than cloud-dependent systems.
These experiments require no budget commitment and can be completed within one week.
Objective: Assess PGC's existing data assets for AI readiness.
Method: Catalog available defect imagery, measurement logs, installation checklists, and technical documentation. Identify gaps in digital preservation versus institutional knowledge held by senior technicians.
Deliverable: A prioritized inventory of domain data assets with a roadmap for digitization and structuring.
Why it matters: This determines whether PGC has the raw material to build competitive AI skills or needs to invest in data capture first.
Objective: Validate whether existing technical documentation can be converted into executable AI skills.
Method: Use an open-source document parsing tool to extract structured content from 2-3 standard PGC installation procedures. Test whether a current skill framework (Claude Code or DeepSeek Harness) can execute queries against this extracted knowledge.
Deliverable: A working prototype of a glazing procedure skill with sample queries demonstrating capability and limitations.
Why it matters: Proves the concept before committing resources to full documentation conversion.
Objective: Determine whether current field hardware can support local AI inference.
Method: Select one representative field device (e.g., tablet used by installation crew). Deploy a lightweight open-source inference engine (such as h3.c or equivalent). Run a benchmark task relevant to glazing (e.g., defect classification from a sample image set). Measure response time, accuracy, and battery impact.
Deliverable: A hardware performance report with clear go/no-go guidance for edge AI deployment.
Why it matters: Edge inference is central to field deployment strategy. This experiment validates whether current hardware investment is sufficient or requires refresh.
The window is open but narrowing. The 12-18 month strategic window identified in prior briefings remains valid. However, ecosystem consolidation is accelerating. The risk is not technological—tools exist and work—but competitive. Early movers in domain-specific skill libraries will establish the standards and capture the market for glazing AI applications.
Domain expertise is the moat. PGC cannot outspend major tech companies on AI research, nor should it try. The defensible position is glazing-specific knowledge: defect recognition patterns, measurement protocols, installation sequences, regional building code variations. Converting this expertise into verified, provenance-protected skills creates barriers competitors cannot easily replicate.
Build for portability, not platform lock-in. No current skill framework is guaranteed to survive. PGC should architect skill development with abstraction layers that allow migration between platforms. Treat skills as assets, not applications.
Edge over cloud for field operations. Cloud dependency creates connectivity risk, latency, and security exposure. Given DRAM scrambling research and verification requirements, on-premise inference should be the default architecture for any field-deployed AI capability.
Monitor the Chinese AI ecosystem. DeepSeek's plugin ecosystem is producing domain-specific skills at a pace Western analysts undercount. PGC should establish lightweight monitoring for glazing-adjacent developments in that ecosystem, including potential partnership or licensing opportunities.
End of Briefing