Sarp Ocali
Founder & CEO
System architecture, core algorithms & product direction
B.Sc. Mechanical Engineering, Politecnico di Torino.
The Gammatic Agent The AI that runs on one principle:
“Measured actions overpower reasonable decisions.”
Every action the AI takes and every assumption the user brings to a still-undefined problem is turned into mathematical functions. Those functions are measured deterministically at every step. That is how Gammatic catches drift, looping, fixation and false finishes while the work is still in progress. On your own infrastructure.
Working with Aurvionics.
You do not need a perfectly defined problem to begin.
Bring the work you want to understand better. We define the problem with you, install Gammatic inside your boundary, and stay to monitor and support it.
Define togetherUnderstand the workDepartment, people, workflow, data, constraints.The task, its checks and its completion criteria are calibrated around the stable core.
Deploy on-premiseIntegrate inside your boundaryBuilt and installed inside your systems and your data boundary.Nothing leaves.
Monitor and supportEvaluate and refineRepresentative tasks are evaluated with your team.Aurvionics monitors and supports the system.
Molded to your problem.
Not the other way round.
Aurvionics does not ship a generic assistant. We start with your department, your files and formats, your systems and your access limits, and define the problem together. The core stays the same; the context, the checks and the completion criteria are calibrated to your work.
We respect your data privacy.
We pick the AI model that fits your job, from design to engineering, from manufacturing to legal and research. We calibrate it to your work and install it inside your boundary, on your infrastructure, including isolated networks. Nothing about your work is sent to an external AI provider. The integration scope is agreed with you, not assumed.
How it works
We don’t ask the AI whether it is on track.We measure it.
The Gammatic algorithm turns every AI action and every user assumption into mathematical functions. As the work progresses, each new prompt and each result is turned into functions again. Those functions are measured deterministically against the objective and its constraints.
The model interprets and proposes. A separate control layer decides what happens next: continue, redirect or stop. The reference is mathematics, not the model’s confidence.
This separates intelligence from authority, the way every safety-relevant machine separates operation from oversight. Whether the evidence is sufficient, whether the direction still holds and whether the work is finished are decided outside the model.
Question the premise. Measure the direction.
Human assumptions and AI actions enter the same loop. A wrong premise from the person is measured the same way as a wrong step from the AI.
Neither the user nor the AI starts with the full picture.
Work rarely arrives as a problem statement. It arrives as an intuition: partial information, a suspected cause, a deadline. Today’s AI produces a plausible next step, and plausible is not verified. Unmeasured, an investigation fails in four recognisable ways, and each of them looks like progress while it is happening.
01 · Direction
Are we solving the right problem?
Drift
Each step moves a little further from the objective, and the next step builds on it.Nobody decided to change course.The work simply did.
Control response
Direction is measured at every step.The next one is redirected before more work builds on a wrong assumption.
02 · Progress
Is this step adding evidence?
Looping
Tools run, text accumulates, the same ground is searched again.More activity is not more evidence.
Control response
Repetition is detected.The approach changes instead of repeating.
03 · Fixation
Are we both stuck on the first explanation?
Fixation
The person suspects a cause.The AI agrees, because agreeing is the plausible move.A shared assumption can go untested for hours.
Control response
The premise is questioned and the alternatives stay open.The missing test is asked for.
04 · Completion
Have we done what was agreed?
False finish
“Done” is declared before the criteria are met.The gap is found by whoever relies on the result next.
Control response
The criteria are checked and what is still open is named.Completion is a governed decision, not a declaration.
One loop for the person.
One loop for the model.
Agree on the goal. Act. Check against measurements. Decide: continue, redirect or stop. You can revise the goal at every turn, and Gammatic questions the framing too, not only the AI’s output.
When the work departs from the objective, the deviation is measured before it compounds. A corrective instruction redirects the next action. The search settles back on task, like a damped oscillation.
Research and engineering in the same room.
Gammatic began as a mechanical engineering thesis at Politecnico di Torino, defended in July 2026 with full marks. It grew into a company with the people who pushed it further.
Goksun Gokce
Co-Founder & CTO
Software development, integration & system validation
B.Sc. Computer ScienceM.Sc. Data Science and Engineering, Politecnico di Torino
Aksel Kemahlioglu
Co-Founder & COO
Operations, project delivery & business development
Interdisciplinary architect, specialised in facade and structural designMaster of Science, Brandenburg University of Technology Cottbus-Senftenberg
Prof. Alessandro Chiaraviglio
Manufacturing & Production Engineering Consultant
Adjunct Professor, Politecnico di Torino
As the project’s supervising professor, he has guided Gammatic from its earliest research stages at Politecnico di Torino.
Prof. Umut Sezen
Software Engineering Consultant
Ph.D., University of Warwick
Advises on Gammatic’s software architecture and development, drawing on experience in engineering and academic research.
Alimurtaza Rutci
Technical Consultant · Systems
M.Sc. in Mechanical Engineering, Istanbul Technical University
Advises on system requirements and industrial integration, helping shape Gammatic around engineering workflows and operational constraints.
Let’s begin with your problem.
Tell us what the work is, who does it, and where the data has to stay. Every message lands in one inbox, read by the founder.