AI troubleshooting for hardware engineers
Be the one who solves the issue nobody else could.
Load the manuals for your tool and Cala works the issue with you. It goes through every kind of root cause, not just the one you were already chasing, and quotes the number for each candidate from the page it came from. It runs on your computer.
Cala is a desktop app that helps hardware engineers find the root cause of a machine problem. Start with the manuals for one tool, not your whole library. Cala reads them into a model of that machine on your computer, then works the problem with you: every kind of root cause, what each check would rule out, and which page every number came from.
01 · The situation
The tool is down, and everyone is looking at you.
Nobody holds a machine this size in their head. Not you, and not the engineer who has been on it since it was installed. What they have is a sense of where to look first.
Without it you start swapping parts. You know it is a guess. So does the person watching you do it.
02 · The loop
Send it up, and what comes back is a list.
Escalate, and you know what comes back. More things to check, or a request for more data. Almost never a reasoned root cause. And the reply takes a day if you are lucky, sometimes weeks, while the customer asks what changed.
Cala gives you ranked hypotheses in about a minute, each with its causal chain written out and the page it stands on. The loop runs at your speed after that, not the queue's.
03 · What you get
The cause hides two subsystems away.
On a complex tool, everything touches everything. A drift in one subsystem gets absorbed by the next and shows up as a symptom in a third, and every box checked on its own comes back healthy. That crossing is what makes the hard ones hard, even for engineers who know the machine.
Cala reasons in causal chains. It works from how the machine actually operates, across every kind of cause at once, and follows the physics across subsystem boundaries no one person owns. What you get is the chain written end to end, in minutes, with the evidence for every link.
One causal chain, three subsystems
A setting drifts out of its window
A downstream loop quietly compensates
The symptom appears, two systems away
Checking each box on its own shows three healthy subsystems. The issue only exists in the arrows.
04 · See it work
One issue. Candidates across every kind of cause.
Here is one issue on a tool. Cala does not jump at the first likely answer. It works through every kind of root cause the machine has, and each candidate names the page it came from and what to go and measure next.
Load lock B reaches transfer pressure 38 s late, but only from the third wafer of a lot onward. Load lock A on the same tool is unaffected. No alarm, no error code.
Kinds of root cause
Candidates · Thermal
Every kind above is a different thing that can go wrong, not a different way of saying the same thing. Cross one off and the problem is smaller.
05 · Check it yourself
Every number comes with the page it came from.
Values are shown the way they appear in your document, with the document name and the page. You can go and check it yourself. Cala can be wrong about which cause it is. It does not invent the number, and if the number is not in anything you have loaded, it says the evidence is missing instead of filling the gap.
That matters most with the customer in the room. Their tool is down and they are counting the hours. Telling them you are going to try another part sounds like guessing, because it is. Walking them through what you ruled out, and the line in their own manual that ruled it out, does not.
Base pressure interlock: 2.4e-2 Torr, alarm at 3 s dwell
VTM service manual, rev C · p. 84
The gauge was zeroed after the last PM. The record does not say.
06 · The other ending
Some nights it ends differently.
The causal chain is on the screen in minutes, the check comes back positive, and the room hears what it was and why. The tool comes back up, and the night ends at a reasonable hour.
That speed is not a shortcut. It is the same first-principles work a great engineer does over weeks, done in front of you in minutes, with the evidence and the checks in order. The investigations on this site are real ones, and some of them had been open for weeks before they were closed in a sitting.
07 · Hand it over
Hand it over with the citations still attached.
When you have to hand it over, write it out: a service report for the ticket, a summary for the next shift, or a hand-over for the customer. Send it up with what you checked, what each check ruled out, and the page you read it from.
Each investigation keeps its own symptom, evidence, ranked root causes, assumptions and next steps. Your notes and measurements sit with them. Next quarter, when the same tool does something similar, you are not working it from a blank page again.
- Symptom
- as written in the ticket
- Checked
- what was measured, in order
- Ruled out
- each with the page that ruled it out
- Assumptions
- labelled, not folded in
- Next step
- ordered by what it rules out
Cala's answers and extracted knowledge are AI-generated. Verify against the source documents and OEM procedures before acting.
08 · Where things live
Your documents stay on your computer.
Cala runs on your computer, and your documents live there. When Cala reasons, the pages being read pass through our proxy to the model provider and come straight back. Nothing is stored on our servers.
- The documents you load
- Extracted knowledge
- Chats, investigations, reports
- Email, plan, usage counters
- Your credit balance
- Nothing else
No document content, and no chat content, on our side.
09 · Questions
The questions that come first.
In the words they usually arrive in, because those are the ones worth answering. The long versions live on the security and privacy pages, and the proof lives in the case studies.
Can it actually solve something my whole team is stuck on?
It has. A machine down two weeks, every sensor in spec, no alarm, no trend. Cala’s top-ranked hypothesis was a software compensation term nobody had questioned; reading the configured value took minutes, and it was roughly ten times its intended value. Restored, the machine requalified.
First hypothesis on the list, confirmed on the first check. Two more investigations are in the case studies.
What does it do when I hand it a real problem?
It works the problem, not the chat. Every kind of possible cause on the table: thermal, mechanical, electrical, contamination, control and software, alignment and metrology. Hypotheses ranked. Evidence quoted with document and page, read facts separated from inferences.
Checks are ordered by what each one rules out, so a quick check that kills three hypotheses outranks a slow one that confirms a favorite. It ends in a report file you can attach to the ticket as is.
Nobody has ever seen my problem. There is nothing to search.
That happened. A dimensional oscillation in dense sections with no write-up anywhere. Cala proposed a material-dependent mechanism nobody had put together, designed experiments to separate it from the alternatives, and attached a prediction to each arm. The experiments ran. Every prediction held.
When there is no page to find, it reasons from the physics in your documents.
It will just tell me what the manual already says.
The manual is where it starts, not where it stops. One process model had been trusted for years, but the fits would not settle on a single parameter set. Cala asked for the one thing nobody had put on the table, the model implementation, and found a quadratic where the physics calls for an exponential. Corrected, every structure collapsed onto one parameter set.
It has no loyalty to what everyone already believes.
It’ll make up a part number. They all do.
Don’t trust it. The product isn’t built on the idea that you will. Every value that decides anything carries a citation naming the document and the page, quoted as written. Inference is labeled inferred. Missing evidence is called missing, not covered with a number of the right shape.
Test it the cheapest way there is: load a manual you know cold and try to break it. There is a ready-made protocol for exactly that.
I have search, and I have ChatGPT.
Search returns what you thought to ask for. A chatbot answers the pages you pasted. Cala tells you what else it could be, and it keeps the machine: documentation read once, a knowledge base for the whole tool on your disk, there every session after. You are not re-uploading PDFs at 3 a.m. on shift.
My documents are OM and PM manuals, BKMs and vendor training decks, half in Chinese.
Load them as they are. That pile of vendor PDFs for one tool is exactly the input shape: Cala reads it into a knowledge base for that machine, mixed Chinese and English without being told which is which, and answers in whichever language you ask. Citations name the document and the page either way.
First users are semiconductor equipment and process engineers. Any machine with real documentation works the same way.
What leaves my computer?
Your documents, and everything Cala learns from them, stay on your computer. Our servers hold your email, plan and usage: no document content, no chat content, and we train nothing on any of it. While Cala is reasoning, your question and the pages being read pass to the model provider through a stateless proxy that keeps nothing on the way. Full detail on the security page.
Do I need my company’s permission?
That is between you and your employer’s rules about software and documents. Cala is a personal subscription, not a way around anyone’s policy. If the conversation needs to happen, the security page says exactly what leaves your computer and what never does, and it was written to be shown to the person who approves things.
A hundred a month is real money, and it’s my money.
20, 100 or 200 USD a month, sized by how much troubleshooting you run, plus prepaid credits when a bad week uses the plan up. Personal subscription on your own card, no procurement. Cancel from your account page any time; it stops at the end of the paid period, and everything on your disk stays yours. Read a case study first and decide from that.
What do I need to run it?
macOS 13 or later, Apple silicon or Intel. Windows 10 or 11, 64-bit; the Windows build is a preview and not signed yet. Internet only while Cala is working a problem. The knowledge base is local and stays local.
Bring the one that's still open.
Load the manual for the tool it is on, and give Cala the issue still sitting there with no root cause. Read the list. Check the citations against the pages they point at. See whether there is something on it nobody has tried yet.