Private AI for the work you can’t afford to get wrong. It runs on your own machines, gives the same answer to the same question, and shows where each result came from.
In daily useat a national research centre in Crete
Weeks of lab worknow done in minutes
60,000+ Supreme Court rulingsin Novit, our assistant for Greek lawyers
35.34° N · 25.14° ENight shift --:--:--
LogCatch 041logged · released
Log · this shift
CatchTimeWhereThen
It fishes inside its own harbour, casts the same way every time, writes down every catch and lets every fish go.
Scroll
01 / 03Private
It runs on your own machines.
It works inside your harbour. Your data never goes out to someone else’s cloud.
02 / 03Consistent
Same question, same answer.
Every cast lands in the same ring. Ask again tomorrow and you get the same result.
03 / 03Sourced
It shows where each answer came from.
Every catch goes in the log: which data, which step, which rule.
01/The problem
Most AI makes you choose between private and powerful.
The most capable AI runs in someone else’s cloud, so using it means sending them your data. The private kind, on your own machines, is usually the weaker one. Most teams end up giving up one or the other.
We build AI that stays on your machines without falling behind. It works best on the repeatable processes your work depends on, once we’ve written down how they’re done.
Businesses now buy AI by the seat, at about $20 a month, and more and more of their use goes to the cheaper, faster models: the top models’ share of that use fell from 53% in August 2026 to 45% by early September (Ramp AI Index, 9 Sep 2026). That is the class Omicron belongs to. The difference is what it is allowed to see: because it runs on your machines, it can read the whole file.
The usual choice
Privacy030
Performance070
Their cloudYour machines
Your data stays with you.You don’t give up performance for it.
02/How it works
AI that follows your method, every time.
Ask a chatbot the same question twice and it may take two different approaches. That’s fine for a first draft. It’s a problem when the answer is a measurement, a deadline, or a number someone has to sign off.
So we don’t let the AI decide how the work gets done. We sit down with the people who do it and write down their steps, settings and rules. Each step becomes a tool that works the same way every time. The AI only picks the tool and fills in the details. If a request breaks the rules, it’s stopped before it runs. In our October test, 22 of 30 rule-breaking requests were stopped before they ran; we are closing the 8 that got through.
Same question, same answer. Run it again tomorrow and you get the same number, to the precision we report: 100 of 100 runs of the tool on the published data, 100 of 100 on a second researcher’s data after a fix on 18 Oct (98 of 100 before it), and through the chat the tool’s one value in 93 of 94 completed runs.
Sourced
Every result shows where it came from: which data, which step, which rule. The request demo on this page is the example; a real receipt from a lab run is written and waits for a read before it goes up.
Tested
We score it against answers you already trust, on cases it has never seen. You get the score with the system.
When a better open model appears, we swap it in and your tools stay as they are.
03/The test
We test it against the AI tools people already pay for, and publish every run.
Each test uses one real input, the same plain-English request for every tool, and many runs, so a lucky answer can’t pass as a result. The full runs are on the proof page.
First test, May 2026
We took one real measurement from the lab and gave it to our tool and to three leading AI platforms, all on paid plans with every feature turned on. Each one got the same file and the same request, four times.
PlatformRunsUsableSame answerWhat happened
Our toolour tool
1234
–/ 4–/ 4
The same result in all four runs, 34 seconds each.
GPT-5.5
1234
–/ 4–/ 4
Saw one peak in some runs and two in others. Its results moved by up to 54.8% from run to run.
Claude Sonnet 4.6
1234
–/ 4–/ 4
Measured the wrong feature in most runs, and drew its boundaries differently each time.
Gemini 3.1 Pro
1234
–/ 4–/ 4
Got a unit conversion wrong every time, so none of its results were usable.
The same kind of test on today’s tools, with many more runs, and a second instrument. On today’s tools the picture is the same as in May: the cloud models gave a different exponent almost every run; our tool gave one.
Runs
Usable
Same answer
What happened
Our tool aloneno language model in the loop
100
100
100 of 100
One value every time, 5 s per run. On a second researcher’s file: 100 of 100 after a fix on 18 Oct (98 of 100 before it).
Omicron with our toolthrough the chat
100
94
93 of 94
The tool returned one value; the model wrote it to four decimals in 72 runs and rounded it in 21. 6 runs stopped before the fit; 1 run gave another value while our own fault test was hitting the same service.
Claude Opus 5.5agent with a shell, paid plan
10
10
4 different values
0.988–1.003; wrote and ran its own fitting code each time.
Claude Sonnet 5.5agent with a shell, paid plan
30
30
19 different values
0.97–1.03; wrote and ran its own fitting code each time.
GPT-6 Lunavendor’s command-line tool, high reasoning
Runs made 17–18 October 2026, through each vendor’s own command-line tool (Codex CLI 0.163.0) or as an agent on a paid plan; the setting is printed on every row; none of these runs used web search. Our runs: 100 per task. Where a vendor’s usage limit cut a run short, the table says so.
Second instrumentElectron-microscope image, dominant period: our typed function gave the same value in both reference runs; GPT-6 Luna (default) gave 9 different values in 30 runs, 26 inside the accepted range and one answering 57 nm; Claude Sonnet 5.5 (agent) 11 different values, 28 inside.
Privacy test, October 2026
When the data may not leave, the model that stays is the one that can answer.
Many files can’t be sent to an outside AI service as they are. A client’s identity is covered by a lawyer’s duty of secrecy, personal data by the GDPR, a supplier’s prices by a trade-secret clause. So we wrote three realistic files (a civil case, a lab measurement campaign, a supply contract), invented from start to finish, and hid the parts those rules protect. Each tool got 20 questions per file. The outside tools saw the file with those parts hidden; Omicron, running on our machines, saw all of it.
We compared against the model each vendor sells most of. On the legal file we also ran each vendor’s top model, so nobody can say we picked a weak opponent.
File
Tool, file with protected parts hidden
Omicron, whole file
Tool, whole file (check only)
Civil case
hidden-field questions 0 of 12 for every tool, all said so · open 8: Claude Sonnet 5.5 100 %, GPT-6 Luna 90 % at default / 94 % at high
97 %
Sonnet 100 %, Luna 81 % / 98 %, Claude Opus 5.5 100 %, GPT-6 Sol 100 %
Lab campaign
hidden-field 0 of 12 for every tool · open 8: Sonnet 100 %, Luna 42 % / 98 %
100 %
Sonnet 100 %, Luna 62 % / 98 %
Supply contract
hidden-field 0 of 12 for every tool · open 8: Sonnet 100 %, Luna 75 % / 99 %
100 %
Sonnet 100 %, Luna 58 % / 98 %
Correct answers out of 20, with the margin of error. “Couldn’t say” counts separately from “wrong”, because saying “this can’t be determined from what I was given” is the honest answer to a hidden field. The whole-file column is a check that the questions are fair; it uses invented data only, and no real file was sent anywhere.
These tests measure consistency, and what each tool can answer with the data it is allowed to see. Choosing the right approach is still the expert’s job, so we build theirs in.
FORTHought is the AI platform we built, in daily use at a national research centre in Crete. Researchers use it every day to search their papers, analyse data and process results from their instruments. It runs on the lab’s own servers, and nothing they send leaves the building.
Spectrum analysis~2 weeksminutes
per batch of measurements · closely matches the researcher’s own results
checked against calibration · used by three researchers
16people outside our core team typed 8,716 messages in 26 weeks, 20 Apr – 18 Oct 2026 · 4 to 9 of them in any one week · counted 19 Oct 2026
BLaw
Novit, a case-file assistant for Greek lawyers
Novit reads the case file and helps with the work around it.
01Works out the right procedure and the right court, following the rules.
02Calculates deadlines and explains each one. If it can’t calculate one, it says so and why.
03Finds Supreme Court rulings for and against your client, most relevant first.
04Drafts the court filing and marks missing information in red, instead of making it up.
Case files are processed on our own machines; nothing is sent to an AI provider.
0+Supreme Court (Άρειος Πάγος) rulings
~10codes, laws and EU acts, up to date as of Law 5326/2026
338amendments tracked, so it reads each article as it stood on the date that matters
6court-filing templates
Figures as of 9 Nov 2026, read from Novit’s live status.
Built on it
A small team built these four apps on one private system, with Omicron, the private model we run on our own hardware, at its centre.
Lab assistant
~2 weeks → minutes, per batch of spectra
Researchers ask in plain words; the instrument’s tool runs their method and returns the numbers with where they came from.
Novit
60,000+ Supreme Court rulings
It reads a case file, works out each deadline with the rule behind it, and drafts the filing with the gaps marked.
Phonē
answers aloud in Greek or English
You talk to it in Greek or English, and it answers aloud in the language you used.
Work, Blackboard, Minutes
A workspace for longer jobs, a board you draw on that the AI can read, and meeting minutes with every speaker named.
It can be optimised for Greek-language work. Two of these already run in Greek today: Novit, on Greek law and Greek court rulings, and Phonē, which listens and answers in Greek.
CNext
Your process
Anything that has to run the same way every time, on data that should stay in-house. Lab analysis, case files, compliance checks, technical reviews.
Four steps from the first call to a tested system.
01 /
Listen
We sit down with the people who do the work and, over a few short sessions, write down how they do it. On our first project, 7 of their 30 answers changed what we had assumed, so we never skip this step.
7 / 30answers that corrected our first version
02 /
Build
We turn each step into a tool that always works the same way. The AI decides when to use each tool, but never changes how the work is done.
03 /
Install
We set it up on your own machines, next to your data and the software you already use. Nothing leaves your building.
04 /
Test
We score it against results you already trust, on cases it has never seen, and show you the number with its margin of error. You see the number before you commit.
After delivery
A member of our team stays on to look after the system and answer your questions. When a better open model appears, you get it, and your method stays the same.
06/Who owns what
What you own, and what we keep.
Yours
the system installed on your machines
your data
your methods and procedures
every result it produces
Ours
The technology we use to build, tune and test your system. It’s what lets us keep improving it after delivery.
07/Who it’s for
Built for teams whose data can’t leave the building, and whose answers have to stand up to checking.
Early exploring, where nobody has settled on a method yet, or tasks where a rough answer is good enough. A general chatbot will do fine there, and we’ll tell you so.
08/Our principles
What you can hold us to.
01
Every number has a source. If we can’t show where a figure came from, we don’t publish it.
02
Research comes first. Our products grow out of our research. When the two pull in different directions, the research wins.
03
We show our evidence. Our claims point to published research and to test runs you can download and check for yourself.
04
We’re upfront about limits. Our paper lists what it doesn’t prove, the proof page does the same, and so will the report you get from us.
05
We don’t train on your data. Nothing you send is used to train an AI model.
09/Our name
Our name
τεκμήριον(tekmírion)τtεeκkμmήíρrιiοoνn
Techmirion comes from τεκμήριον (tekmírion), the ancient Greek word for proof: the evidence that settles a question. We chose it because proof is what we deliver.
Our logo is built on the small filled square that mathematicians put at the end of a proof. You’ll see it on this page wherever we’ve shown ours.
claim
same question, same answer
source
Adamidis et al., 2026, Table 3
checked
4 runs, the same result
The square that closes a proof
τεκμήριον — proof. We’d rather show you the evidence than ask for your trust.
10/Who we are
Built by the researchers who use it.
Techmirion is a company founded by the researchers who built FORTHought, the AI platform in daily use at a national research centre in Crete.
We’re a small team of researchers and engineers. We built FORTHought and the apps on it, Omicron answers every request they make, and we rely on all of it every day. What we offer you is what we already use ourselves.
We work to scientific standards, as an independent company in Heraklion.
Supported by
Prof. Emmanuel Stratakis
11/Research
Read the research behind it.
PreprintMay 2026
Adamidis et al., 2026
Done since the paper
a re-test on today’s AI tools with many more runs (17–18 October 2026)
a second instrument (an electron-microscope pattern)
a privacy test on data that may not leave (three invented files)
thirty harder questions on which the private model scored level with the top paid models
What we’re working on next
01a formal study with users
02support for more lab instruments
03systems at more organisations
12/Next
Systems that improve their own method.
On our own hardware, our model now proposes rewrites of its own tool descriptions and instructions. Every proposal is checked for false facts, tested blind against the unchanged version on questions it has never seen, and kept only if it wins by a clear margin. A kept change is then tested again on a sealed set of questions that nobody tunes on.
The first rounds, in October 2026, kept some changes and rejected others. We are re-testing the kept ones on larger samples before we print a figure.
We think private AI goes this way next: a model on your own machines that gets measurably better at your work, under your control.
What has to happen before a change reaches anyone
01Fact check
02Blind test against the unchanged version
03Re-test on a sealed set
04A person approves
Get in touch
Start with one process.
Pick the one that takes the most time, or would do the most damage if it went wrong. We’ll tell you plainly whether we can help, and what results to expect.
It can be anything your team does the same way every week: a measurement, a deadline, a filing, a review. We write down how you do it, build it, install it on your machines and score it against answers you already trust. You see the score before you decide anything further.
A first project ends with a working system and a written test result, whether or not you continue. What a first project looks like