PhD candidate in Computer Engineering at Hacettepe University · Technical lead of the HÜSMART autonomous shuttle · Co-founder of Binary Brain
I work on representation learning, and on the perception stacks that have to survive contact with the real world. Three things run in parallel right now: a visual-inertial odometry system for a Level-4 shuttle, a JEPA for drug-target interaction, and an AI products company with apps in both stores.
alpdemirel.com · LinkedIn · Medium · X
Flow-aware VIO for dynamic urban scenes · poster at ToRK 2026, paper in preparation
Classical VIO assumes the world holds still. City traffic disagrees. This is a stereo pipeline on the OpenVINS MSCKF estimator that does two things: it masks dynamic objects using instance segmentation gated by ego-motion-compensated optical flow, and it recovers ego-velocity from quasi-static traffic participants, using them as additional filter anchors.
On the VIODE benchmark, ATE improves in 9 of 12 environment/density conditions, by 26 to 67% at the traffic densities that matter operationally. The result I care about most: at high parking-lot density the method beats a ground-truth segmentation oracle, 1.26 m ATE against the oracle's 1.48 m. The oracle dutifully masks every parked car, and parked cars are excellent features. Motion matters, not semantic class.
DTI-JEPA · PhD research, Biological Data Science Lab, advised by Prof. Tunca Doğan
Most drug-target models learn a similarity metric and report numbers on random splits. The question here is whether a joint-embedding predictive architecture can learn protein-to-ligand structure well enough to generalize to targets it has never seen. SaProt-650M (LoRA) encodes structure-aware protein tokens, MolFormer-XL encodes the ligand, and the predictor is conditioned on a 3D pair-distance ALiBi bias taken from the resolved complex. Trained with smooth-L1 against EMA targets plus SIGReg, with no contrastive objective. Evaluated cold-target.
HÜSMART · technical lead
Türkiye's first university-built autonomous shuttle, on a VW ID. Buzz platform, targeting SAE Level 4. ROS 2 Humble and Autoware, four Hesai LiDARs (OT128 / JT128), four cameras, GNSS/INS, NDT localization against Lanelet2 HD maps. I own the technical architecture and the roadmap. Development is simulation-first in AWSIM while the vehicle is in procurement.
| Project | What it is |
|---|---|
I-JEPA from scratchPyTorch ViT SSL |
Does predicting in representation space beat predicting in pixels? Same ViT-B/16, same data, same budget, no timm and no pretrained weights. A frozen linear probe reaches 77.6% on STL-10 against my own MAE at 72.7% with full fine-tuning, which is the harder protocol winning. The README documents the three silent bugs that nearly killed the run. · write-up |
Masked AutoencodersPyTorch ViT SSL |
The pixel-space baseline for the comparison above, built the same way. 75% masking, L1 reconstruction, and a feature space that separates cleanly before a single label is seen. · write-up |
Semantic Anomaly DistillationDINOv2 MobileNetV3 edge |
Traffic accident detection, where a missed crash costs more than a false alarm. Distilling world knowledge from a frozen DINOv2 teacher into a MobileNetV3-Small student moves the decision boundary toward sensitivity: +4 recall over standard training at 100+ FPS on edge hardware. · write-up |
Fighter Jet Survival TacticsUnity ML-Agents DQN A* |
Hybrid A* handles the route, a DQN handles the missile. Five million training steps, a custom PID flight controller, and a runtime switch between deliberative planning and reactive evasion. · write-up |
EKF-RPL for high-mobility networksOMNeT++ C++ Kalman |
RPL routing that predicts where nodes are going and prunes unstable parents before links break. Two lightweight 1D Kalman filters instead of full 2D matrix machinery, for constrained devices. PDR 33.96% to 36.32% under Gauss-Markov mobility at 5 to 15 m/s. |
Council of High Intelligence (Gemini CLI)agents MCP |
Eighteen personas deliberate a hard decision in structured rounds across multiple providers, so the disagreement is real rather than one model arguing with itself. Ported to Gemini CLI, with cross-provider MCP routing added. · write-up |
Also here: an agentic code-generation framework benchmarked against human developers on Pylint and maintainability index, and the OpenVINS fork the VIO work builds on.
I write about the work while it is still uncertain, mostly for Towards AI.
- Beyond Tokens: How JEPA Is Quietly Teaching AI to Understand the World
- I Built I-JEPA From Scratch and It Beat My Own MAE, With a Frozen Encoder
- How I Built a Real-Time Traffic Accident Detector Running at 100+ FPS
- I Built an AI Pilot That Plans Like a Robot and Dodges Like a Human
- I Host My Backend From My Own Laptop for $0
Research · PyTorch, self-supervised learning (I-JEPA, V-JEPA, MAE), vision transformers, knowledge distillation, protein and molecule encoders, reinforcement learning (DQN, PPO)
Robotics · ROS / ROS 2, Autoware, OpenVINS, LiDAR + camera + IMU + GNSS fusion, Lanelet2, NDT, AWSIM and CARLA
Edge and embedded · NVIDIA Jetson (Xavier NX, Nano), CUDA, C/C++, STM32, FPGA, Raspberry Pi, real-time control loops at 1 kHz
Product · Flutter, Unity (C#), FastAPI, Docker, Google Cloud, Oracle Cloud
At Binary Brain I co-built and shipped two Unity apps on iOS and Android: Practice with Me, an IELTS tutor with an AI avatar doing real-time lip-sync over on-device speech-to-text and text-to-speech, and Dream Meow, a generative AI mobile experience. Current work there is Flutter with FastAPI backends.
Before robotics I spent three years asking how a weakly electric fish weighs vision against electrosense, which turns out to be sensor fusion under uncertainty with a 400 million year head start. That became an MSc on multisensory integration, six international conference presentations (FENS, SICB, SfN, SIU), 1 kHz embedded control loops and 25 Hz vision pipelines on a Jetson Xavier NX, and most of my instincts about hardware. Supported by an MSCA fellowship and TÜBİTAK.
If you work on VIO in dynamic scenes, JEPAs outside of vision, or shuttles that have to drive on real streets, I would like to hear from you. Open to research collaborations, and to roles in Europe.