Projects
What I have built, in research and in industry, and what it is evidence of. Each card says plainly whether it is published research, delivered software or a proposal. For the mechanisms behind the research, the Research Lab has twelve short films.
Research
SecurePoL with Watermarking
The reference implementation behind my dissertation and the SecurePoL paper: proof-of-learning verification that also checks a feature-based model watermark, so a forged training record fails.
Adversarial Examples for Proof-of-Learning
Attacks that try to pass proof-of-learning verification without doing the training. Breaking the protocol first is what shaped the defences in the SecurePoL line of work.
Inside a Level D flight simulator
How a certified simulator earns a pilot's trust, and what a learned model would have to keep to run inside its real-time loop. Every rule is cited from the regulation text; the thesis is my own and not yet published.
Cyber-security events from Turkish text streams
My METU master's thesis under Prof. Cengiz Acartürk: detecting rare security events in Turkish Twitter and newspaper data, the basis of our ICISSP 2021 paper.
Industry
The Avion cards describe my role, the architecture and the engineering practice. They disclose no customer, budget or security specifics of the company.
AI strategy and reference architecture
I wrote Avion's AI adoption roadmap and a reference architecture that separates on-premises inference, governed cloud APIs and edge workloads, built the portal that presents it, and built engineering-knowledge retrieval prototypes. The interactive view shows the design, including agents that are proposed rather than deployed.
Software for Level D full-flight simulators
Real-time simulator software in C, C++ and Scala, and a telemetry pipeline that sustains about 50 GB/s through batching and bounded-queue backpressure, for simulators certified at the highest grade. The field note covers what such a platform offers for evaluating learning systems.
Simulator monitoring and diagnostics
Web applications for the Avion Cloud Environment that monitor, configure and diagnose core simulator components (TypeScript, Svelte, Scala, Python, gRPC): time-aligned state of the kind any learned component would have to be evaluated against.
Mission control for autonomous UAVs
Mission-control and ground-control software for the Kargu and Togan UAVs in C++ with Boost and Qt, including power-on and continuous built-in tests, under hard-real-time, safety-critical constraints.
Havelsan DLP
Team lead for Havelsan's data-leakage-prevention product for defence and government clients, and developer of its endpoint agent in C, C++, Java and Go. The same policy-enforcement problem now applies to prompts, retrieval indexes and agent actions.
Secure Web Gateway, Patch Manager and Dragon
Windows services and applications in C++ for Comodo's Dome Secure Web Gateway, enterprise Patch Manager and Chromium-based Dragon browser; I also set up the team's unit testing, code review and Jenkins CI.
Proposed applied-AI workflows
Designs from the Avion AI strategy, written as proposals rather than deployments. Each keeps a person accountable for the outcome and works only over approved company data; each would need evaluation before use.
- Engineering knowledge copilot. Version-aware retrieval over engineering documents, with source citations and the specifications a change may affect.
- Repository agent. Proposes review findings and maps changes to requirements in self-hosted pipelines; release approval stays with the team.
- RFP accelerator. Drafts cited answers from approved past responses and product documentation, for experts to verify.
- Instructor debrief writer. Drafts a training debrief from telemetry and instructor notes, with student identifiers removed; the instructor owns the assessment.
- Fault diagnostics copilot. Retrieves similar past cases and bulletins to suggest where to look next; it never clears equipment for service.
- Qualification lifecycle copilot. Links requirements to test evidence when drafting qualification plans; it is not a substitute for qualification results.
- Predictive maintenance forecaster. Estimates component wear from telemetry, if reliable failure labels exist and it beats the current maintenance rules on held-out data.
Other
EKG measurement and annotation tool
A calibrated image-measurement utility built and tested with cardiology professionals. It is a manual tool, not a diagnostic model; measurements of this kind are what a computer-vision model would be checked against.