Open-source AI, deployed against the factory floor.
Field guides written from a practitioner's workstation, not from a vendor's slide deck:
anomaly detection that trains on good parts only, model-to-edge toolchains
(ONNX → RKNN on RK3588-class NPUs), synthetic data generation for lines that
never produce enough defects, and the camera-to-model ops in between. Every entry is
grounded in what the workstation's disks can prove — versions, commands, artifacts — and
states its honest boundary when a line was staged but not yet flown. The rubric is
always what it is → what it does → how to run it → what went wrong.
Numbers land in the /data/ ledger.
A live manufacturing plant: frame-by-frame vision inspection, three-a-day patrol agents, and a 27B on-prem LLM — the first unsupervised night shift, 2026-09-05.
VERIFIED · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · on-prem machine-room workstation with a single RTX 4090D 24GB serving a Qwen-family 27B model, 4-bit quantized, via vLLM #computer vision#industrial inspection#AI agents#local LLM#vLLM
Unattended dress rehearsal before an AI agent's first night shift: five gates on 2026-09-05 — 25-s ignition check, zero-second alignment, canary, clean restore.
VERIFIED · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · rehearsal executed on the live lines over the encrypted read-only collection path; all timestamps local plant time, logged #AI agents#industrial inspection#night shift#guardrails#sentinels
The rehearsal's most precise moment — two lines, same second, zero error at 19:00:00 — came from one line of Bash arithmetic, not a model.
VERIFIED · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · the rehearsal and the overnight shift run the same frozen script on the line-side containers; all timestamps local plant time, logged #determinism#AI engineering#guardrails#industrial automation#LLM ops
An on-prem 27B's hardest day: two blank papers, one config race, zero model swaps — four real production days passed after the exam rules changed.
VERIFIED · Changeover-daily pipeline (LineWatch family) on two live production lines at a discrete-manufacturing plant · on-prem Qwen-family 27B, 4-bit quantized, on one RTX 4090D 24GB under vLLM · incident 2026-09-10 night, fix and re-exam 2026-09-11 #local LLM#vLLM#incident review#LLM ops#industrial automation
Token audit of the plant's only LLM call: ~52K steady-state tokens per call, ~97% door fee, zero tools used — rebuilt same-model to ~2K in one day.
VERIFIED · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · audit covers the AI-summary lane's full production life 2026-09-05→09-11 (44 calls, transcript-measured) · rebuild and A/B on 2026-09-12, same model both sides #LLM ops#AI agents#token economics#guardrails#industrial automation
Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses.
VERIFIED · LineWatch patrol system on two live production lines at a discrete-manufacturing plant · staircase assembled 2026-09-12 from the Episode-4 audit and the same-day lane rebuild · dual-lane comparison mornings 2026-09-11→09-13 (12 windows, two recipes) · ruling delivered 2026-09-13 · the rebuilt lane's first live event shift 2026-09-12 20:55, with the next morning's green shift as the zero-call control #LLM ops#AI agents#token economics#industrial automation#manufacturing
The Episode-4 audit came home to the operator's own desk: 1.8B tokens in 30 days, 95% of it cache re-reads. The cure — compaction — costs detail memory.
VERIFIED · The operator's home fleet — a three-node private mesh (cloud services node + on-prem GPU workstation) · all evidence gathered on-box on 2026-09-12: thirty days of interactive coding-agent session transcripts aggregated per-session from usage fields, the chat gateway's own ledger read from its local store (2,564 metered calls since 2026-07-25), compaction applied the same day to the gateway's four largest sessions and verified at the wallet level #LLM ops#AI agents#token economics#context windows#prompt caching
anomalib on a two-GPU workstation: all 15 MVTecAD scenes trained and measured (mean image-AUROC 0.981), plus the three real bugs the sweep had to fix first.
What breaks (GitHub, HF, PyPI, Omniverse — with receipts), what doesn't (domestic inference), and the architecture that worked: rent a computer, not a tunnel.
VERIFIED · Editorial synthesis — no new measurements; every wall and route cited is first-party evidence from this site's published dispatches (2026-08-26 to 08-30) #networking#Tailscale#China AI#mirrors#edge deployment
The synthetic-data plan, three staged repos, 19 zero-length launch logs, and the download wall that stopped it — a field note on a line not yet flown.
VERIFIED · Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) — RTX is the entry ticket this line never got to use #synthetic data#Isaac Sim#Omniverse#ROS#domain randomization
PatchCore from the workstation, byte-identical checkpoints on a Mac mini, a Basler camera through Aravis, 0.156-0.176 s CPU inference — a deployment that ran.
VERIFIED · Deploy side: Mac mini, Apple Silicon, CPU inference (anomalib 2.0.0, torch 2.7.1, Python 3.10.18) · Train side: RTX 4090D/A4000 workstation (anomalib 2.1.0.dev0) · Basler GigE camera via Aravis #edge deployment#anomalib#PatchCore#Mac mini#Apple Silicon
Meta's SAM3 on one workstation: a scripted docker door never pulled, a 131 MB source clone, and a real 533-line ComfyUI node package with saved blueprints.