Mehmet Kerem Turkcan

I build AI systems that have to work outside the lab, where latency and reliability matter as much as accuracy. My research covers urban perception, medical robotics, machine learning systems, and computational neuroscience. I release much of it for others to build on: real-time detectors, open datasets and models, and research platforms. I also make independent games and films.

At the Center for Smart Streetscapes (CS3) and Civil Engineering & Engineering Mechanics at Columbia University, I lead machine learning projects from research prototype to deployed system. I created DART for real-time open-vocabulary detection and led UrbanOmniDetect and UrbanOmniView for calibration-free monocular 3D perception. For pedestrian and cyclist safety, I built bikeped and contributed to PAVE, and I led a city-scale study of congestion pricing using 910 public traffic cameras. This work also includes street-scale data pipelines, synthetic data generation, urban digital twins, and vision-language models on edge devices.

In medical robotics, I work on surgical world models, open datasets such as Open-H-Embodiment, and Surgical SAM 3.1 for segmenting instruments and anatomy. This builds on my postdoc work with Northwell Health collaborators, where I applied computer vision to robotic surgery and endoscopy training.

DART Open-vocabulary detection at deployment speed UrbanOmniDetect View-agnostic urban detection Medical Robotics World models and open datasets for medical robots

Work and leadership

• Fielded AI systems: I lead multimodal perception projects whose sensing models are built to scale across 900+ New York City intersections and CS3’s three urban testbeds: COSMOS PAWR in New York City, DataCity in New Brunswick, and FAU MobIntel in West Palm Beach.
• Publications: My papers appear at CVPR, ICML, ACM UIST, ACM/IEEE SEC, IEEE INFOCOM, IEEE PerCom, and EDM, and in IJCV, eLife, and Surgical Endoscopy. Our paper on real-time video analytics for urban safety received the Best Paper Award at SEC 2025.
• Open-source systems: DART and generative-agents, two of my public AI systems, have 1,200+ GitHub stars combined.
• Sponsored research: I have contributed proposals, technical reports, sponsor reviews, and engineering to about $30M in research funded by NSF, DARPA, AFOSR, and Con Edison, with additional project support from NVIDIA and EmpireAI.
• Teaching and mentorship: I taught graduate deep learning courses at Columbia, mentor Master’s and high school researchers, and was the main engineering instructor for the CS3 Research Experience for Teachers in 2024, 2025, and 2026.

My work on machine learning systems includes Loom, an analytical neural computer that runs compiled C programs inside a looped transformer. With collaborators, I have also worked on adaptive data collection for robust learning, vision-language models split between cloud and edge, and the security of edge-cloud systems. Earlier, I wrote GPTune, a GPT-2 fine-tuning toolkit from the early public LLM era.

Before my current role, I was a postdoc at the AIDL Lab in Columbia’s Department of Electrical Engineering, where I also earned my Ph.D. My doctoral research was on the fruit fly brain: I co-designed and built FlyBrainLab, an open platform for exploring its circuits and an early prototype of today’s AI research workbenches. Its natural-language interface linked papers, ontologies, and a connectome knowledge graph to GPU simulation and interactive 3D visualization.

My game and film work with Wisedawn and KEDIKAT also feeds my research: experience with real-time engines and visual storytelling shaped simulation projects such as Boundless. KEDIKAT’s film FLEA received four festival accolades in 2026, including Best Storytelling Runner-Up at the MetaMorph AI Award.