“Show every reject with a wrinkle near the seal since Tuesday.”
Filters the reject archive by defect class, image region and time, then returns the matching frames with timestamps and line positions.
Vision & Inspection
Ask your inspection history a plain question
A private vision-language copilot on Gatepoint G8 lets operators and quality engineers search rejects, explain decisions and draft shift reports in plain language — on site, with no cloud, while your inspection models keep pass/fail authority.
The problem
Thousands of reject images, and no time to look
A modern inspection system keeps an image of every reject. That archive should answer the questions a quality engineer asks: do seal wrinkles cluster after a film change, which camera has been uncertain all week? In practice, answering means exporting records and scrolling a gallery for an afternoon. Shift handovers get written from memory.
Vision-language models (VLMs) read images and text together, which suits exactly this kind of search and summary. They can also be confidently wrong. And for plants that keep production images on site, a cloud chat assistant is not an option.
So we built the copilot around one rule: it can help people find, understand and report — it cannot decide whether a part ships.
How it works
A hybrid architecture with clear authority
Two kinds of model run side by side on your hardware. The split between them is fixed in the design, not left to configuration.
| Role | Deterministic inspection models | Vision-language copilot |
|---|---|---|
| Pass/fail decision | Sole authority, through validated reject logic | None |
| Speed | Under 40 ms frame-to-I/O on the line | Seconds per answer; nobody is waiting on a reject gate |
| Output | Verdict, defect class and confidence | Answers with cited frames, summaries and draft reports |
| Changes to reject logic | Only through your change-control process | Impossible — read-only access, no path to PLC or I/O |
| Where it runs | Gatepoint at the line | Gatepoint G8, on site, no internet required |
In practice
Questions operators and engineers ask
Every answer comes back with the frames and records it is based on, so the person asking can check it in one click.
“Why was the pack at 2:32 p.m. rejected?”
Explains the inspection model's verdict from its class, confidence and heat map, and links the exact frame and model version.
“Did rejects change after the 2 p.m. film changeover?”
Compares reject counts and classes either side of the changeover logged in the MES, with example frames.
“Which camera has the most borderline calls this week?”
Ranks cameras by low-confidence decisions and suggests frames for the retraining queue — for a person to approve.
“Draft the night-shift quality summary.”
Drafts reject counts, top classes, notable frames and open questions for the shift lead to edit and sign.
“Does anything on camera 3 look unusual right now?”
Describes the live scene and flags anything unfamiliar for a person to review. It never rejects a part itself.
Guardrails
Built so a wrong answer can't do damage
It cites its frames
Every answer links the frames, records and timestamps it relied on. A claim without a source is not shown as fact.
It logs everything
Each question, answer and cited frame is written to the Perflens Console audit log with the user and model version.
It can't change reject logic
The copilot runs under a read-only service account with no write path to the PLC, the I/O or the model registry.
It says when it finds nothing
If the search returns no matching records, the copilot reports that instead of composing an answer.
Access follows your roles
Single sign-on through SAML 2.0 or OpenID Connect, with role-based access so each user sees the lines they work on.
It stays on site
Model weights, the search index and the images all live on the Gatepoint G8 and your servers. No internet connection is required.
Process
From a list of questions to a working copilot
-
Question workshop
We collect the questions your teams spend the most time answering, and write down what a correct answer looks like for each.
-
Connect and index
A Gatepoint G8 connects to your reject archive, live feeds and MES events, and builds a searchable index on site.
-
Measured pilot with one shift
The copilot is tested against a written set of questions with known answers. We report how often answers were correct and how often citations pointed to the right frames.
-
Rollout and care
More users and lines, operator training, and model updates delivered through Perflens Care with the same evaluation set rerun after each update.
Configuration
Typical copilot configuration
- Hardware
- Gatepoint G8, NVIDIA Jetson Thor-class
- Data sources
- Console reject archive and defect gallery, live camera feeds, shift and changeover events from your MES over OPC UA or MQTT
- Model
- Open-weight vision-language model sized for on-device inference; no external API calls
- Authority
- Read-only; pass/fail stays with the deterministic inspection models
- Access and audit
- SSO via SAML 2.0 or OpenID Connect, role-based access, every prompt and answer logged
- Network
- No internet connection required
Honest fit
Who the copilot is for
A good fit when
- You already have an inspection image archive and too little time to mine it
- Quality engineers spend hours on reject analysis and customer complaint investigations
- Several lines and shifts need consistent handover reports
- Images must stay on site, or the plant runs without internet access
Probably not when
- You want an AI that decides pass or fail on its own — this one is designed never to
- There is no image history yet; start with AI visual inspection
- One low-volume camera whose defect gallery a person can review in minutes
See it in context
Where a copilot earns its place
Heat-seal inspection at 180 packs per minute
A modeled snack line whose reject archive a copilot can search by defect, region and shift.
Read more
Food & beverage
Fast lines, frequent changeovers and handovers that matter.
Read more
Where VLMs help and where they hallucinate
Mara Lindqvist on putting vision-language models next to a production line, and keeping them out of the decision.
Read moreFAQ
Questions we hear often
Can the copilot pass or reject a part?
No. It has no write access to reject logic, PLC outputs or the model registry. Pass/fail decisions stay with the deterministic inspection models and the validated logic around them.
What happens when it gets an answer wrong?
It can, which is why every answer cites its frames and records. Before rollout we measure accuracy on a written set of questions, and that evaluation is rerun after every model update.
Does it need the internet or a cloud AI service?
No. The model, the index and the images stay on the Gatepoint G8 and your on-premise servers. Read more in air-gapped AI.
Does it work with inspection systems Perflens didn't build?
Often, if the existing system can export reject images with timestamps and verdicts. We check this during discovery; the copilot is easiest to add where the Perflens Console already holds the archive.
Do we need new hardware on every line?
No. Your existing Gatepoints keep running inspection. A single Gatepoint G8 typically serves the copilot for several lines, depending on camera count and how many people use it.
Bring the question that takes an afternoon to answer
Tell us what your quality team spends hours searching for. We'll tell you whether a copilot answers it — or whether a better report would.