I build the systems that let cameras see, decide, and adapt in real time.
From GStreamer pipelines on edge devices to orchestration over frontier vision-language models, my work turns live video into decisions at city scale. I am Co-founder and CTO of Cyrus Technology, after a Ph.D. in the WiMNet Lab at Columbia University with Prof. Gil Zussman.
Feedback loop: analytics quality feeds adaptation, which feeds the next frame
Research interests
Systems for seeing in real time
My work sits where computer vision meets distributed systems. The question I keep returning to is how to keep analytics accurate and fast when the environment, the network, and the hardware all change underneath you. That means building the full stack: GStreamer and DeepStream pipelines in C/C++ on NVIDIA Jetson devices, controllers that pick stream configurations online, and edge-cloud architectures that combine CNN detectors with large vision-language models.
real-time video analyticsedge/cloud computingmultimodal AI & VLMsreal-time system designonline optimizationwireless & wired networks
Selected work
Projects
Ph.D. thesis core · 2019 – 2025Patent filed
ASTRA: adaptive streaming for real-time video analytics
Video analytics breaks quietly when light, traffic, or network conditions drift. ASTRA is a low-overhead runtime that watches analytics quality and retunes each camera stream’s configuration online, using a constrained extension of GP-UCB to stay near-optimal across multi-camera pipelines. Presented at ACM/IEEE SEC’26 and filed as a US patent.
A deployed system that extracts and fuses traffic and crowd information from bird’s-eye and street-level cameras, running across edge and end devices at a live New York intersection.
A hierarchical, distributed hybrid twin for urban traffic management that joins AI, edge-cloud computing, and next-generation networks. Won IDC’s 2023 Smart Cities North America Award for transportation.
StreetNav: street cameras guiding blind pedestrians
Street cameras localize pedestrians far more precisely than GPS. StreetNav uses that signal to give blind pedestrians turn-by-turn outdoor navigation, built with Brian Smith’s group at Columbia.
Pipelines in C/C++ and GStreamer with YOLO detectors and the NVIDIA DCF tracker that extract and fuse traffic and crowd information from two viewpoints, demonstrated live at PerCom and MobiSys.
An automated C/C++ and GStreamer pipeline that measures distancing compliance from street cameras, evaluated in the COSMOS testbed and published in the IEEE Internet of Things Journal.
EdgeCloudAI: distributed video analytics with VLMs
An asynchronous, event-driven framework in C/C++ with NVIDIA GStreamer plugins that splits work between CNNs on the edge and large vision-language models in the cloud.
Cyrus Technology is an AI video analytics company in San Francisco. We build an orchestration layer over frontier vision-language models for real-time analysis of camera feeds, and a generative AI copilot for live video surveillance. The company won the NSF-sponsored CS3 lab-to-market accelerator and grew out of the camera-based security work I started at Columbia’s Center for Smart Streetscapes, where the project won the CS3 Ignition Award.
Lead an AI video analytics company that builds an orchestration layer over frontier vision-language models for real-time analysis of camera feeds.
Led development of a generative AI copilot for real-time video surveillance.
Won the NSF-sponsored CS3 lab-to-market accelerator.
Jun. – Aug. 2024
PhD Founder Track Summer Program
Participant
Completed a six-week program for PhD founders on commercializing research, scaling projects, and raising first capital.
Feb. – Sept. 2019
Electrical Engineer
NOVE · Tehran, Iran
Designed and implemented a C/C++ algorithm that de-interleaves radar signals and identifies pulse repetition frequency (PRF).
Classified radar signals (constant, jittered, staggered) from time-of-arrival analysis, combining an improved sequential difference histogram with an improved PRI transform.
Grouped received pulses by emitter and determined each emitter’s PRI, pattern, type, and source location.
Built a Qt interface that visualizes de-interleaved radar signals and their parameters in real time.
Research experience · WiMNet Lab, Columbia University
Research
Fall 2024 – Dec. 2025
Distributed Video Analytics Systems over Edge and Cloud
Designed distributed architectures that integrate CNNs and large vision-language models for high-performance video analysis.
Built an asynchronous, event-driven, ultra-low-latency framework in C/C++ with NVIDIA GStreamer plugins.
Fall 2019 – Dec. 2025
ASTRA: Online Edge-Cloud Video Analytics Adaptation Patent filed
Developed ASTRA, a low-overhead runtime architecture that adapts camera stream configurations in real time to maintain analytics performance under unknown environment changes.
Designed a controller that selects near-optimal configurations for multi-camera pipelines, extending GP-UCB to constrained settings.
Presented ASTRA at ACM/IEEE SEC’26 and filed a US patent application (Pub. No. US 2026/0163998 A1).
Fall 2022 – Dec. 2025
Real-Time Multi-Camera Analytics for Traffic and Crowd Information
Developed real-time pipelines that extract and fuse traffic and crowd information from a bird’s-eye-view camera and a street-level camera (C/C++, GStreamer, YOLOv4–v8, NVIDIA DCF tracker).
G. Jain, B. Hindi, Z. Zhang, K. Srinivasula, M. Xie, M. Ghasemi, D. Weiner, S. A. Paris, X. Xu, M. Malcolm, M. Turkcan, J. Ghaderi, Z. Kostic, G. Zussman, and B. Smith
G. Jain, B. Hindi, M. Xie, Z. Zhang, K. Srinivasula, M. Ghasemi, D. Weiner, X. Xu, S. A. Paris, C. Tedjo, J. Bassin, M. Malcolm, M. Turkcan, J. Ghaderi, Z. Kostic, G. Zussman, and B. Smith
Poster description in Proc. ACM ASSETS’23, Oct. 2023
