Portfolio tooling nudges

Data scientists are getting a sharper portfolio brief: ship one Generative Artificial Intelligence project tied to a business problem, make the workflow rerunnable, and show the result in a lightweight app. (x.com) Matt Dancho’s July 2026 post argued that a portfolio should center on one project with a clear business use case, not a stack of disconnected demos. He paired that with a requirement that the work be reproducible, so another person can run the same steps again instead of relying on screenshots or one-off notebooks. (x.com) A separate GitHub Projects post pointed readers to Violit as one way to package that work. The framework’s GitHub page describes it as a pure-Python web framework with “fine-grained” updates instead of full-script reruns, and its Python Package Index release 0.4.2 was published on April 5, 2026. (x.com) (github.com) (pypi.org) Violit’s pitch is simple for portfolio builders: write Python, get a dashboard, and avoid some of the plumbing that slows down app demos. Its documentation says apps can be built in a single Python file, with more than 70 widgets, more than 20 themes, and partial updates instead of rerunning the whole script after each interaction. (doc.violit.cloud) (violit.cloud) The process advice in the third post came from Alexey Grigorev, who mapped Cross-Industry Standard Process for Data Mining, or CRISP-DM, onto Artificial Intelligence system design. His framing treated prompt design as part of the modeling phase, placing prompts alongside model choices as something teams test and revise rather than treat as a final polish step. (x.com) (developers.openai.com) (docs.cloud.google.com) CRISP-DM is older than the Generative Artificial Intelligence boom, but it was built for the same management problem: making analytical work repeatable across teams. The original CRISP-DM guide, published in 2000 by a consortium including SPSS, NCR, and DaimlerChrysler, defined six phases — business understanding, data understanding, data preparation, modeling, evaluation, and deployment. (public.dhe.ibm.com) That older framework explicitly aimed to make projects “more repeatable” and less dependent on one expert operator. A 1999 paper on the method said the goal was a process that different people could run reliably and adapt to different situations, which matches the current push for rerunnable notebooks, pipelines, and app demos in Artificial Intelligence portfolios. (cs.unibo.it) Prompt design fits that structure because model behavior changes with the instructions it receives. OpenAI describes prompt engineering as writing instructions that consistently produce the required output, and Google Cloud says the work is iterative, with teams repeatedly updating prompts and assessing responses. (developers.openai.com) (docs.cloud.google.com) The combined message from the posts is narrower than “build more projects.” Build one project that starts with a business question, runs end to end more than once, and is easy for another person to inspect in a small application instead of a static slide deck. (x.com 1) (x.com 2) (x.com 3)