Application Refactoring · IT Operations Automation
NVIDIA needed AI tooling that could autonomously handle complex engineering sessions and full machine-learning research workflows, not just short coding assists, while respecting privacy and infrastructure constraints.
Reported impact
NVIDIA accelerated both software delivery and ML experimentation, lowered the threshold for building internal tools, and gave researchers a more autonomous partner for exploring ideas, tracing evidence, and launching end-to-end studies.
NVIDIA uses Codex with GPT-5.5 as a default tool for complex engineering work and end-to-end machine-learning research. In software delivery, Codex can sustain long sessions with multiple compactions, choose the right tools or skills, and evolve internal systems from MVP to production-ready services with better scalability and reliability than earlier models. Engineers have used it to build and autonomously test a podcast-recording app in hours, including video and audio workflows via the Codex desktop app. In research, the same model acts as a creative partner: teams point it at large corpora of papers, let it trace evidence chains, map knowledge graphs, write experimental scripts, and run workloads remotely through SSH on machine-learning infrastructure. That combination means Codex can move from idea generation to experiment execution and system implementation in one flow, including rewriting legacy Python repositories into far more efficient Rust when performance improvements justify it.
Engineering tickets, internal platform code, and research corpora are fed into Codex, which reasons with GPT-5.5 to propose designs, generate code, run tests, and launch remote experiments through SSH. Outputs are then reviewed or executed on NVIDIA’s infrastructure as production systems or ML research runs.
"Codex is our go-to tool for complex engineering tasks"
"It’s been a 10x speed improvement just in terms of running experiments"
"it was built and tested completely autonomously"
"it’s rewriting it into Rust and making it like 20X more efficient"
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
OpenAI Sample Oct 9, 2026 Asana Asana needed to cut the cost and latency of its StackAI browser agent so it could offer customers more capable models for browser-driven workflows without unsustainable operating expense. $0.47 average estimated model cost with the optimized GPT-6.1 Sol workflow Codex · Command · GPT-6 Astra OpenAI Sample Oct 8, 2026 Braintrust Braintrust wanted to respond to customer feature requests faster and run more engineering experiments, but slower coding tools and backlog-driven workflows made it costly to turn feedback into working prototypes in real… 50% of the Braintrust team moved to Codex in one month Codex · GPT-5.5 OpenAI Sample Sep 23, 2026 Cognition Cognition is the company behind Devin, an autonomous software engineer used by businesses around the world, from big banks to tech-native startups. As Cognition’s engineering teams write more code, reviewing that work… Cognition is applying GPT‑6 Astra across its product lineup. "We’ve been using it to make improvements across… GPT-6 Astra 