By Travis Oliphant, William Martinez-Torres, and Emma Jo Merrill
Generative AI has introduced a new paradigm for rapid software development, often described as “vibe coding” [1, 2]. While this approach accelerates prototyping, it can also rapidly increase technical and architectural debt as it can lead to significant deficits in reliability, maintainability, and security. The growth in “Vibe Coding” creates a growing gap between prototypes and production systems [3, 4, 12, 14].
This challenge is actually analogous to the state of scientific computing in the 1990s, before the maturation of the scientific Python ecosystem [5]. This paper posits that the principles that led to Python’s success—in particular the development of a robust, community-driven ecosystem of hardened, interoperable libraries, and the integration of Python with existing frameworks built with other languages [8, 9]—provides a necessary model for production-ready AI.
We believe a “Guided Vibe” paradigm will be necessary, where AI-assistance is contextualized within a structured platform. This approach shifts the developer’s task from generating unconstrained code to composing validated, high-level components, such as is available in the Nebari OS AI ecosystem [10]. This method synthesizes the velocity of AI-assistance with the engineering discipline required for scalable, enterprise-grade systems using a secure but open source platform.
Generative AI tools are enabling a new development method, “vibe coding,” characterized by natural language-driven code generation [1, 2]. The primary advantage of this method is decreased latency between people with the most knowledge and understanding of what to build and code that provides a working version. Using natural language as a direct interface to constructing software, allows non-developers to quickly iterate and experiment to produce a prototype.
Despite this initial velocity, a “prototype-to-production gap” is becoming evident. Industry reports from 2024-2025 indicate a high failure rate for AI pilots (up to 40%) moving to production [3, 4]. Analysis of AI-generated code reveals inherent weaknesses: prototypes are often brittle, difficult to maintain, and contain critical security vulnerabilities [12, 13]. Academic studies confirm that iterative AI generation without expert human oversight can increase security flaws and introduce structural deficits not found in human-written code [12, 14]. Furthermore, this problem is aggravated during full-stack development, where keeping the moving parts of development in unison is marred when the AI model is stretched thin across the whole stack.
This situation is not novel. It mirrors the challenges faced by the scientific Python community in its nascent stages [5]. The solution, then as now, can be found in establishing a mature ecosystem of robust, reusable components. We believe a “Guided Vibe” approach is the future, leveraging the historical lessons of Python’s “librarification” [8, 9] to create an open source framework for reliable AI-assisted development within a structured platform like Nebari OS which includes the benefits of peer review, community-driven improvements, and therefore, community-ownership [10].
Python’s initial adoption in scientific and engineering domains was driven by its simplicity, accessibility, and collaborative open source nature. It functioned as “executable pseudocode” [6, 7], making it easier for domain experts to implement and test algorithms [7] without becoming software engineers. Generative AI tools are enabling a new development method, “vibe coding,” characterized by natural language-driven code generation [1, 2]. The primary advantage of this method is decreased latency between people with the most knowledge and understanding of what to build and code that provides a working version. Using natural language as a direct interface to constructing software, allows non-developers to quickly iterate and experiment to produce a prototype.
The community’s need to validate, share, and scale computations exposed the limitations of this approach. Reproducibility was a significant challenge, as scripts that functioned in one environment would fail in another [15]. This demonstrated a clear need for standardized, performant, and reliable tools that could serve as a common foundation for development.
Python’s successful transition from a scripting language to a dominant production ecosystem for data was not inevitable; it was a direct result of “librarification.” This process involved encapsulating complex, high-performance logic (often written in C/C++/FORTRAN) into stable, high-level Python APIs. The creation of NumPy, in particular, provided the foundational n-dimensional array object, a robust component that abstracted away an entire class of complex, error-prone memory and loop management [8]. SciPy, built atop this foundation, provided fundamental algorithms for scientific computing along with a plethora of other libraries, methods, and distributions [9]. Developers moved from writing everything to composing solutions from hardened, trusted components as a community, sharing lessons learned.
The availability of these standard components (NumPy, SciPy, Pandas, Scikit Learn, and the Jupyter project) fostered a powerful ecosystem effect. It created a lingua franca for data and science, enabling interoperability between different tools and, critically, fostering a community that could build and share best practices [16, 17]. The lesson for modern AI development is clear: sustainable velocity and reliability emerge from a common foundation of robust, composable building blocks, and a sense of community-driven ownership: not from ad-hoc generation of low-level code.
Crucially, Python succeeded not just because it was easy to write, but because it created pathways for integration and constant improvements. It became the glue that connected disparate systems such as high-performance Fortran/C/C++ code and databases, deep learning frameworks like PyTorch and TensorFlow, and rich UI frameworks like Bokeh, Enaml, and Panel. This ability to integrate existing, proven technologies into a cohesive whole is a core tenet of building production-ready systems. “Guided Vibe” extends this principle, combining AI with community-driven efforts to intelligently stitch together these powerful, existing frameworks rather than attempting to reinvent them from scratch.
We posit that a “Guided Vibe” paradigm, which applies the lessons of Python’s maturation to AI-assisted development, is necessary for future development of the applied AI ecosystem. This model reframes and augments the role of the AI. Instead of a generator of arbitrary code, the AI acts as an intelligent integrator and navigator. The developer’s task becomes one of high-level composition, using natural language to select, configure, and connect trusted, pre-validated components within a specific, known context, not just writing code, but bringing enduring solutions.
This paradigm is entirely dependent on context. The AI’s suggestions must be constrained by the “guardrails” of a defined platform, which includes a curated set of available libraries, established architectural patterns, and security policies. For example, within an AI-assisted notebook, a request to “load data” would not generate a novel parsing script; it would instead generate the correct, idiomatic usage of pandas.read_csv (as specified by the platform’s trusted component list), followed by a call to a known data validation library. The resulting workflow is maintainable, transparent, and built on production-ready components from its inception.
A “Guided Vibe” requires a practical, curated ecosystem to provide the necessary context. Nebari OS, an open-source AI platform, emerging from our previous work on Nebari for data-science and ML/AI deployment is evolving to serve this role [10]. We are developing Nebari OS in an open-source pattern with broad input from the community which includes the emerging applied AI society. NebariOS is one example of an approach to provide a curated connection to the “guardrails” that make AI-assisted development safe and scalable through several key features:
The Nebari OS platform’s configuration is the context. It informs and constrains the AI’s suggestions, guiding development toward patterns and components that are known to be scalable, secure, and maintainable within its architecture.
To operationalize this paradigm, we have partnered with the emerging Applied AI Society (https://appliedaisociety.org) and have introduced the “Applied AI Rodeo” [11]—a hands-on workshop methodology. These workshops (“wrangling buzzwords into real-life applications” [11]) function to train developers in this component-first, Guided Vibe methodology. Facilitated by AI Architects who are seeded from our OSA (Open Source Architect) community, participants build and deploy production-aware solutions using the Nebari OS ecosystem, demonstrating the synthesis of AI-driven speed and component-based reliability.
We are observing a familiar pattern in the industry. The current “pure vibe” approach to AI development offers rapid prototyping but creates systems that can be brittle, insecure, and difficult to scale [3, 4, 12, 14]. This mirrors the state of scientific computing before the “librarification” of Python, where the community learned that sustainable development relies not just on writing code, but on building upon a foundation of hardened, community-vetted libraries[8, 9, 17]. We believe the industry is naturally trending toward a “Guided Vibe” paradigm, where AI is used to compose trusted components rather than generate infrastructure from scratch. We are evolving our open-source AI operating system, Nebari OS, to build on a history of open-source innovation to provide the stable scaffolding for this approach. We cannot build the future alone. We invite the community of builders, maintainers, and architects to join us in defining what “Guided Vibe” means, helping us synthesize the velocity of AI with the engineering rigor required for the next generation of software.
[1] Karpathy, A. (2025, February). Post on X. As cited in Cloudflare (2025), “What is vibe coding? | AI coding.”
[2] Google Cloud. (2025). “Vibe Coding Explained: Tools and Guides.” Google Cloud Blog.
[3] S&P Global Market Intelligence. (2025, March 14). “AI project failure rates are on the rise: report.” As reported by CIO Dive.
[4] WorkOS. (2025). “Why Most Enterprise AI Projects Fail — and the Patterns That Actually Work.” workos.com.
[5] Millman, K. J., & Aivazis, M. (2011). “Python for Scientists and Engineers.” Computing in Science & Engineering, 13(2), 9-12.
[6] Oliphant, T. E. (2007). “Python for Scientific Computing.” Computing in Science & Engineering, 9(3), 10-20.
[7] Lin, J. (n.d.). “A Hands-On Introduction to Using Python in the Atmospheric and Oceanic Sciences.” (Self-published educational material).
[8] Harris, C. R., Millman, K. J., van der Walt, S. J., et al. (2020). “Array programming with NumPy.” Nature, 585, 357–362.
[9] Virtanen, P., Gommers, R., Oliphant, T. E., et al. (2020). “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python.” Nature Methods, 17, 261–272.
[10] Nebari Development Team. (n.d.). “Nebari: Your open source data science platform.” nebari.dev.
[11] National Association of Supplemental Benefits Insurance (NASBI). (2025). “2025 Supplemental Benefits Forum: AI Rodeo: Wrangling Buzzwords into Real-Life Applications.” nasbi.org.
[12] Abbassi, H., et al. (2025). “Is LLM-Generated Code More Maintainable & Reliable than Human-Written Code?” arXiv preprint.
[13] Cloudflare. (2025). “What is vibe coding? | AI coding.” Cloudflare Learning Center.
[14] Alshammari, M., et al. (2025). “Security Degradation in Iterative AI Code Generation: A Systematic Analysis of the Paradox.” arXiv preprint.
[15] Perez, F., & Granger, B. E. (2007). “IPython: A System for Interactive Scientific Computing.” Computing in Science & Engineering, 9(3), 21-29.
[16] Scientific Python. (n.d.). “About.” scientific-python.org.
[17] Bressert, C. (2012). “SciPy and NumPy: An Overview for Developers.” O’Reilly Media.
[18] Quansight. (n.d.). “Nebari Services.” quansight.com.
Like many data practitioners, I have a love-hate relationship with SQL. I like how it’s mostly standardised, portable, popular, and lets you express complex logic. But I also dislike how it can sometimes feel clunky and hard to read.
