Your full-stack integration partner
Edge ai Solutions
Acting as an extension of your engineering team, CTai LABS transforms NVIDIA’s developer ecosystem into production-ready, field-validated solutions.
CTai LABS turns NVIDIA’s robotics and Edge ai technology into rapid deployable reality; enabling customers to move from prototype to production faster, smarter, and with greater reliability.
Built around Connect Tech hardware, CTai LABS delivers an end-to-end service model that combines ai integration, embedded systems engineering, firmware development, model optimization, and edge deployment all under one roof.
Featured Article
We Budgeted 64GB.
Super Mode Said Otherwise.
Two generative models on one NVIDIA Jetson 32GB module:
How Super Mode and Agentic AI changed the sizing equation
JetPack 7.2 MAXN_SUPER brings Jetson AGX Orin 32GB close to the compute class of the 64GB module for workloads that fit in memory. One engineer supervising an AI agent migrated a complete dual-model video AI stack in days, then benchmarked every step. Here is the full engineering story.
By Doruk Sonmez, M.Sc.
AI Solutions Architect, CTai LABS, a department of Connect Tech Inc.
NVIDIA DLI Certified Instructor
Published July 2026 | Benchmark data measured on device by CTai LABS | Technical review: Rob Callaghan, P.Eng., Chief Product Officer, ConnectTech.com
Key Results
- NVIDIA Cosmos open physical AI world foundation model throughput up 65 percent at the production request shape, with roughly 6x faster time to first token
- Optimization improves idle free RAM up from 5.3 GB to about 10 GB, enough room to add an entire second generative model to the same module
- Full NVIDIA JetPack 6 to JetPack 7.2 migration, DeepStream 9.0 port, and a 30-row benchmark campaign completed in days
- Three live 720p30 RTSP streams sustained with continuous alerting on a single NVIDIA Jetson AGX Orin 32GB module
In production Edge AI, the model is only one part of the challenge. Integration is where the system succeeds or fails. A new operating environment, updated inference runtimes, hardware video pipelines, and a fixed memory budget all have to work together before a single frame can be analyzed reliably.
When NVIDIA released JetPack 7.2 with new agentic development tools, memory efficiency guidance, and MAXN_SUPER for Jetson AGX Orin 32GB, we put those capabilities to the test on our Scene Analyzer Agent, a real-time video analysis application built around the the NVIDIA Metropolis Blueprint for Video Search and Summarization (VSS). The target: run the entire stack, both generative models included, on one Jetson AGX Orin 32GB module. (A workload we had previously scoped for a 64GB configuration.)
This is the engineering story behind that migration: what moved, how an AI coding agent compressed weeks of platform bring-up into days, and exactly which memory levers made the 32GB module carry the load. Every number below was verified against live device telemetry, and we report the trade-offs as measured, including where the envelope ends.
Figure 1. Production throughput at 2048 input tokens, 256 output tokens, concurrency 4, across all seven Cosmos serving configurations. Cosmos-Reason2-2B FP8 static kv8 on vLLM leads at 177.0 tokens per second.
Go-to-Market-Faster
CTai LABS Simplifies Physical AI So You Can Build Smarter, Faster
Deployment Support
Generative AI for Edge Compute Applications
Jetson Application Enablement
Processor & Platform Integration
Sensor Integration
-
Integration Scope: Full-stack ai, plus hardware and software integration and/or multi-sensor, thermal compliance
-
Jetson Support: Full Orin/Thor line, custom carrier support
-
Sensor Support: cameras, IMUs, LiDAR, stereo, radar, custom & redundant sensor arrays
-
Software Enablement: bring-up and demo models, Isaac ROS, TensorRT support, custom app tuning, continuous integratation, dev, and deploment
-
Delivery: dev-ready prototypes, app-ready system image, field deployable and integration support
NVIDIA Expertise
NVIDIA Elite Partner
CTai LABS takes NVIDIA’s most advanced tools and frameworks and brings them into real-world deployment:
- Isaac ROS leveraging accelerated ROS 2 packages for perception, navigation, sensor fusion, and hardware abstraction across Jetson and IGX platforms.
- Isaac Sim using NVIDIA Omniverse-based simulation to generate synthetic data, validate perception pipelines, and refine robot behavior before deployment.
- Metropolis Microservices enabling scalable video analytics, real-time monitoring, and multi-camera ai applications at the edge and across distributed networks.
NVIDIA Jetson Thor and IGX Thor the new generation of ai compute designed for robotics, healthcare, industrial, and safety-critical systems.
Our Core Values
