// teknofest 2026

Ten competitions in one year, ten separate fields

In my first year at university I entered ten different TEKNOFEST competitions. Most of them were not my field: nuclear reactor design, underground mining, airline scheduling. Each one meant learning the domain first, then working out where computation actually helps.

Looking back, the projects turn out to say the same thing. The nuclear engine, so the calculation stops depending on imported software. The mining platform, running in a browser because there is no connectivity underground. The language model squeezed onto a Raspberry Pi so it never has to reach the cloud. The video analysis kept entirely local so no data leaves the building. The wallet shield, built to survive a post-quantum world. All of them are after the same thing: systems that work without depending on someone else's cloud, licence or connection.

None of this was assigned. I could, so I did and honestly, the fact that it sounded cool was part of it :D

10competitions
07as captain
09teams founded
03entered alone
02national finals

The deepest runs

Two national finals; the third stopped at the last review before the final.

Nuclear energy · Neutronics·team of 3·captain

Aegis-Q

Nuclear Energy Technologies Design Competition

🏁 National finalist

Problem

The software that computes safety margins for a small modular reactor — criticality, cooling adequacy — is both expensive and entirely imported. The most critical step of the design depends on someone else's licence.

Approach

A computation engine simulating an SMR's neutronic and thermal-hydraulic behaviour, with a 3D digital-twin interface on top. The engine was validated against the IAEA-2D benchmark problem.

What I took away

A computation engine is worth nothing for the numbers it produces, only for its ability to defend them against a known reference problem. Without the IAEA-2D benchmark we would have had plausible-looking output — which, in nuclear, means nothing at all.

Competition page
Blockchain · Cryptography·team of 3·team member

Q-ADAPTIVE

Blockchain Competition

🏁 National finalist

Problem

"Harvest Now, Decrypt Later": attackers are not trying to break today's encrypted traffic, they are storing it — to open retroactively once quantum hardware matures. A key that is safe today leaks the past tomorrow.

Approach

An autonomous key-rotation shield for ERC-4337 smart wallets: on-device AI anomaly detection, an off-chain ZK-STARK proof, an on-chain key swap. Because the key rotates on a provable schedule, the window a leaked key opens stays narrow.

What I took away

The real architectural question is not which algorithm to pick but where each step runs. On-chain is expensive and public, off-chain is cheap but needs proof, on-device is private but unverifiable. The ZK-STARK is precisely the bridge between the three.

Competition page
Mining · Graph neural networks·solo·captain

Q-Twin WebOS

Mining Technologies Competition

Pre-final roundEliminated at the online review that precedes the final.

A first-time competition, entered alone.

Problem

Underground monitoring produces point alarms: a threshold was crossed in one tunnel, a siren went off. Nobody sees how neighbouring tunnels affect one another. On top of that, the geological data — drilling, geochemistry, magnetics — sits in separate places.

Approach

A decision-support platform that runs in the browser and works with no connection. A graph neural network propagates risk spatially across tunnels and produces a risk-aware evacuation route; the scattered geological layers merge into a single view.

What I took away

There is no connectivity underground — and that is not a footnote, it is the constraint that decides the whole architecture. "Make it work offline too" bolted on afterwards produces patchwork; given up front, it changes everything from the data model to the interface.

Competition page

Past the middle rounds

Both were technically standing; one fell on the report, the other in the exam room.

Robotics · ROS 2 · Embedded·team of 9·team member

Otonom Endüstriyel Mobil Robot

Industrial Robotics Applications Competition

Pre-final roundThe report scored low overall and we were cut before the final.

The software and control layer was the highest-scoring section of the report: 8.33 / 10.

Problem

Round-the-clock autonomous pallet handling in narrow warehouse aisles: follow the guide line, verify the station by QR, talk to factory doors and PLC systems.

Approach

A dual-processor architecture — a Jetson Xavier NX for high-level decisions, SLAM and navigation (ROS 2 Humble, SLAM Toolbox, Nav2/TEB), and an STM32 running FreeRTOS for motor PID and sensors. A 250 ms delay in the LiDAR stream came down below 50 ms through the DDS QoS profile. Line following via HSV moments in OpenCV, 3D QR pose estimation via SolvePnP; a checksum-validated UART protocol, a Modbus TCP/IP bridge to the factory PLC, a PyQt5 operator panel and a Socket.IO progressive web app.

What I took away

Real-time behaviour is a budget problem, not a library problem. Those 250 milliseconds were not won by making code faster but by dropping to the right layer and relaxing the transport's delivery guarantee. Finding which layer held the problem took longer than fixing it.

Competition page
Multimodal AI · Turkish NLP·team of 4·captain

Multimodal Video Analiz ve Karar Sistemi

Turkish AI Language Agents Competition · Scenario 3

Technical examSystem failures during the technical exam ended the run.

Problem

Autonomous monitoring of live camera streams across public, security and industrial networks by an AI agent. The binding constraint: data privacy forbids any external API — everything runs on local servers, with no connection.

Approach

Periodic frame capture from the RTSP stream, a vision-language model served locally through vLLM, and a multi-agent pipeline: one agent reading the imagery, another producing a Turkish summary and timestamped JSON. An operator decision panel on top.

What I took away

We picked the hardest of the three scenarios and I would do it again — but here is what I took away: a system that works and a system that works in the exam room are not the same thing. Ours ran fine on our own machines. Delivery turns out to be part of the engineering.

Competition page

Stopped at the first round

Five of the ten ended in round one. Each still left something behind one of them started the field I work in today.

Logistics · Optimisation·team of 2·captain

PRISM

AI-Assisted Logistics Trunk-Line Optimisation Competition

A planning core that folds demand forecasting and route/vehicle assignment into a single engine.

What I took away

The hard part of optimisation is not wiring up the solver, it is deciding what to optimise. Write the cost function wrong and you will find the wrong answer very efficiently.

First round

Aviation · Scheduling·team of 2·team member

GNN-SA Hibrit

AI-Assisted Airline Optimisation Competition

Rescheduling departure times in 10–15 minute micro-shifts. A graph neural network locates the connections carrying the most delay risk; simulated annealing then builds the timetable without violating slot, minimum-connect-time or aircraft turnaround constraints.

What I took away

Constraints do not narrow the solution space, they define it. Any timetable produced before the slot and connect-time rules were modelled properly was flawless on paper and unflyable in practice.

First round

Model compression · Edge computing·solo·captain

Q-Compress Lite

TÜBİTAK 2242 University Students Research Project Competition

A hybrid compression method pairing tensor-train decomposition with 4-bit quantisation, aimed at running the model offline on a Raspberry Pi 5 class device.

What I took away

Compression is a trade, and the question is not "how small" but "where are you willing to lose quality". It was the first research proposal I ran entirely alone; keeping scope narrow is what it taught me.

First round

Quantum computing·team of 3·captain

Quantum Technologies Competition

We prepared by working through IBM's quantum computing training programme.

What I took away

The earliest exit on this list left the deepest mark. Linear algebra and the foundations of quantum computing settled here — the quantum machine learning I work on now is a direct continuation of that preparation. Being cut did not end the interest, it started it.

First round

Open source · Linux·solo·captain

Pardus Bug Hunting and Suggestion Competition

Hunting for defects through daily use and reporting findings in the competition's format.

What I took away

The mechanics of contributing to an open source project from the outside: release flow, how Git structures a repository, what makes a bug report actionable. This is where I saw that you can contribute to a project without writing its code.

First round

// where this campaign led

Five of ten attempts ended in the first round. There is no point hiding it: I did not win most of these. What I did get was the habit of reading a problem in an unfamiliar field, understanding its constraints, and putting a defensible proposal together inside two months.

The earliest exit was quantum. The IBM training I took to prepare settled linear algebra and the foundations of quantum computing and the quantum machine learning I focus on now comes straight out of it. The weakest line on the list is the one that brought me where I am.