July 04, 2026
Pacific Glazing Corporation | July 04, 2026
Local AI has crossed from experimental to production-ready, giving PGC a 12-18 month strategic window to deploy domain-specific tools in field operations. The barrier to entry for building proprietary glazing AI has collapsed—but only for firms that invest now in training data quality. Security concerns around self-improving agents are real but manageable with structured safeguards; the bigger risk is inaction.
What it is: On-device inference frameworks (local-llm, Talos, hermex) that run AI models without cloud dependency.
Why it matters: Field operations can now run AI on tablets and phones without internet connectivity or data leaving the job site. PGC gains speed, privacy, and reliability. This is no longer aspirational—it is production-ready today.
What it is: Systems like Kulaxyz/self-learning-skills that let AI agents harvest successful patterns and encode them as reusable rules without human retraining.
Why it matters: A glazing agent could learn from each installation, refining error-checking logic over time. The catch: misaligned reward functions risk "reward hacking." Requires sandboxed execution and explicit goal definitions.
What it is: The attack surface created when AI agents write persistent code to files or databases.
Why it matters: CVE spikes around frontier models signal new vulnerability classes. Local deployment reduces exposure, but structured input schemas and real-time safety monitoring are non-negotiable for any field deployment.
What it is: Native iPhone apps (hermex) controlling self-hosted LLM agents—not chatbots, but AI tools designed for mobile.
Why it matters: Field technicians get AI assistance on their phones without cloud round-trips. This directly enables PGC's Site Survey Assistant as an iOS app.
What it is: Tools like claude-real-video enabling scene-aware analysis from video feeds.
Why it matters: Automated site surveys become possible. Glass defect detection during installation could catch issues before they become callbacks. Potential to reduce inspection time by 60-80%.
What it is: Mature sustained document analysis supporting comprehensive project specification review.
Why it matters: AI can ingest full project specs, CAD files, and installation protocols simultaneously. Reduces specification review time and catches conflicts before field work begins.
What it is: Techniques to remove specific training data influence from deployed models (e.g., client project data).
Why it matters: GDPR/CCPA compliance requires this when AI trains on client data. Expect unlearning capabilities as a procurement RFP requirement within 12 months. This is not optional for PGC.
What it is: Self-hosted code generation models like Codex-5.5 (1,302 GitHub stars) as alternatives to GitHub Copilot.
Why it matters: PGC can build internal glazing tooling without vendor lock-in or subscription costs. Custom automation scripts, specification generators, and project management tools become affordable.
What it is: Hybrid AI approaches (G-RRM, Program-as-Weights) combining neural networks with symbolic logic for verifiable, explainable decisions.
Why it matters: Structural glass calculations require safety-critical precision. Neuro-symbolic systems provide auditable reasoning—critical for liability documentation and regulatory compliance.
What it is: Peer-to-peer inference networks like Talos enabling GPU rental outside traditional cloud providers.
Why it matters: Inference costs could drop 50-70% versus AWS/GCP. For PGC, this means affordable on-premises or hybrid inference infrastructure with predictable economics.
Local Infrastructure + Mobile Interfaces: The combination of local-first models and mobile-native apps (hermex pattern) creates the foundation for PGC's field AI strategy. No cloud dependency means no latency, no data exposure, and no subscription risk.
Self-Improvement + Security: Agent self-improvement amplifies security risks. Self-improving agents writing persistent code create expanding attack surfaces. Any self-improvement deployment requires sandboxing and schema-constrained inputs.
Video Understanding + Long-Context: Site survey video feeds analyzed alongside project specifications enable real-time comparison. AI watches the install and checks it against the spec simultaneously.
Unlearning + Client Data: Any model trained on proprietary client project data requires unlearning capabilities. This connects GDPR compliance to procurement requirements and competitive differentiation.
Open Code Models + Decentralized GPU: Self-hosted code generation running on cheap GPU infrastructure makes internal tooling economically viable for mid-size firms like PGC.
Duration: 2 days Cost: Under $500 Setup: Install Ollama on existing field tablets. Deploy a 7B parameter open model (Llama 3.1). Connect hermex-style mobile interface.
What to test: Can field technicians run basic glazing Q&A (specification lookup, installation protocol recall) entirely offline? Measure response accuracy and latency on current hardware.
Next step: If latency is under 3 seconds for common queries, proceed to Experiment 2.
Duration: 1 day Cost: Developer time only Setup: Define JSON schema for three common field inputs (measurement submission, defect report, site condition log). Build input validation layer around local LLM.
What to test: Does structured input reduce hallucination rate compared to free-form prompts? Measure error rate on 50 test cases.
Next step: If structured inputs reduce errors by 50%+, mandate this pattern for all field AI tools.
Duration: 3 days Cost: Under $1,000 Setup: Use long-context model to ingest a 200-page glazing specification document. Build prompt library for common extraction tasks (thermal performance requirements, structural load limits, glazing types).
What to test: Can a field technician upload a project spec and get accurate extraction of key parameters in under 60 seconds?
Next step: If accuracy exceeds 90%, this becomes the foundation for automated pre-installation review.
Competitive window: The 12-18 month production-ready window is real. Firms that build training data moats now will have entrenched advantages. Inaction means competing against well-trained glazing AI within 24 months.
Training data is the moat: The strategic asset is not the AI model—it is the proprietary training data (historical projects, CAD files, installation protocols, defect logs). PGC should commission a glazing-specific training dataset immediately. This cannot be outsourced or replicated quickly.
Field-first, not office-first: Mobile-native interfaces and local inference make field deployment the logical starting point. PGC's MVP experiments should focus on field technician tools, not back-office automation.
Security is a prerequisite: Structured inputs, sandboxed execution, and safety monitoring are not optional enhancements. They are prerequisites for any deployment. Budget for this overhead from day one.
Procurement requirements incoming: Within 12 months, unlearning capabilities and security monitoring will appear in client procurement RFPs. PGC should build these in proactively, not reactively.
Cost structure opportunity: Decentralized GPU marketplaces and open code models fundamentally change the economics of AI adoption for mid-size firms. The capital expenditure barrier has collapsed. PGC can build custom tooling without vendor lock-in or enterprise contracts.
End of Briefing