Have you deployed LLMs in air-gapped factory workshops?¶
TL;DR: Yes. Auto-parts workshop air-gapped LLM deployment + Tier-1 line real-time visual QA + Enterprise multi-tenant gateway. All hard KPIs preserved on portfolio.html and PDF.
3 Representative Manufacturing Projects¶
1. Auto-Parts Workshop Air-Gapped AW36-14B¶
- Business problem: workshop air-gapped (physically no external network), 20+ years of expert knowledge in retiring foremen, apprentices had nobody to ask on rare faults, avg incident triage 45min-2h
- Solution: AW36-14B on single edge 4090 in workshop · 20+ years of SOPs + incident archives + key part drawings fed into RAG
- Technical: Ollama single-node · local vector DB · workshop tablet + PC dual endpoints · no public egress · on-site tech regularly syncs new knowledge offline
- KPI: incident triage time -45% · air-gapped fully operational · digital knowledge base for future generations
- My role: hardware selection rationale (why 4090 vs A800: no multi-user concurrency needed) · edge deployment scripts · on-site rollout
2. Tier-1 Auto-Parts Real-Time Visual QA¶
- Business problem: traditional visual QA staff fatigue → defect escape → mounting customer complaints
- Solution: production-line camera real-time visual inference (custom YOLOv8 + defect classifier), decision → direct PLC divert
- Technical: edge inference (industrial PC + 4090) · periodic model retrain · on-site UPS · hot-standby dual-machine
- KPI: defect escape rate -74% · first-pass yield +8% · one-pass rate +12% · 6-month payback
- My role: vision model architecture · line integration · on-site 24×7 SLA
3. Enterprise Multi-Tenant LLM Gateway (incl. Industrial Subsidiaries)¶
- Business problem: large group with 20+ business lines (incl. 3 industrial subsidiaries), varied AI needs, runaway token cost
- Solution: unified LLM gateway → internal AW36-72B (industrial-sensitive) + DeepSeek-V2 (general) + selective public APIs, department token budget caps
- Technical: OpenAI-compatible gateway · Redis metering · department quota · audit log
- KPI: 20+ business lines shared on one stack · per-department token cost capped · group-wide compliance passed
- My role: gateway architecture · token metering · multi-tenant isolation
Key Actions for Workshop Deployment¶
Air-gapped / half-air-gapped delivery is fundamentally different from office LLM deployment:
- Hardware selection: don't blindly buy A800 clusters · single 4090 / RTX 6000 Ada often fits workshop edge better
- Air-gapped availability: all dependencies (model / embedding / vector DB / index / dashboard) shipped once, no external network
- Workshop conditions: dust / temperature / humidity / voltage fluctuation / UPS / hot-standby — all planned front
- On-site ops: workshop tech isn't an AI engineer, dashboard must be minimal, alerts must be direct
- Knowledge refresh: quarterly IT does offline sync of RAG index from HQ
- Level-3 protection / industrial-control compliance: 3-tier data · industrial network · all front-loaded before sizing
What Manufacturing Projects Won't Do¶
- ❌ "Cloud SaaS LLM" for a workshop — the workshop is air-gapped by definition
- ❌ "Cloud annotation platform" — data leaving the workshop = violation
- ❌ "Let GPT-4 analyze your production data" — SLA breaks find no accountable party