An R&D Week at Twentynext: From Research Paper to Production Code

Quick summary

Twentynext’s R&D approach blends scientific research with real-world deployment in the same working week. Instead of keeping research and client work in separate lanes, the two deliberately feed into each other. CRISP-DM acts as the common framework for every iteration, from the initial business question to a deployed model.

• Twentynext gives employees dedicated time for research alongside client work
• Junior data scientists typically contribute to external R&D projects within three months
• The CRISP-DM methodology guides both client projects and internal R&D work
• Collaboration with university medical centers and industry partners shortens the path from paper to practice
• ISO-certified service and maintenance processes help keep models reliable after go-live

From the bench to the lab: how most agencies waste their people (Projects)

A data scientist at the average consultancy usually knows the routine. A project wraps up, the client is happy, and then comes the wait for the next assignment. That in-between period—sometimes days, sometimes weeks—is politely called “bench time.” In reality, it often means momentum stalls, and someone with proven deep learning expertise ends up revisiting online courses they’ve already outgrown.

Twentynext takes a fundamentally different approach. Bench time simply isn’t part of the model. The space that opens up between or alongside client projects is used for R&D: reading recent papers, building prototypes, and testing new techniques on real business challenges with external partners. This isn’t a nice-to-have on the side—it’s part of how the company works.

That has a direct impact on quality. A data scientist who read a paper last week on adaptive color analysis in immunohistochemistry images and applies that technique this week in a proof of concept for digital pathology brings a very different level of depth than someone working purely project to project. Clients benefit from knowledge that is not only current, but already pressure-tested.

The growth of data science in organizations has been driven by more data, better storage, and greater computing power. Even so, most projects still fail to deliver the value expected. Research shows that the vast majority of teams use no formal process model at all. That is exactly the gap Twentynext aims to close—not just for clients, but internally as well.

What you can do:

• Check whether your data science team has explicitly allocated R&D time alongside client work (at least half a day per week is a common minimum)
• Ask your vendor which process model they use; no answer—or “we work agile” without any further explanation—is a red flag
• See whether recent papers and methodologies show up in project documentation; if not, the team’s knowledge base may already be lagging behind

How CRISP-DM connects R&D and client work (Services)

CRISP-DM (Cross-Industry Standard Process for Data Mining) is the most widely used data science process model in the world and forms the backbone of the Twentynext approach. The model consists of six phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. What makes CRISP-DM especially useful is that it is not linear.

Iteration is the point

The phases are not meant to be followed in a rigid sequence; in practice, you often move back and forth between them. In Twentynext’s R&D workflow, that means a data scientist who discovers halfway through modeling that the original business question needs to be reframed simply goes back to phase one. That is not treated as failure. It is how the method is supposed to work.

That is also what sets CRISP-DM apart from a waterfall process. A data mining project does not end at deployment. What you learn can trigger new—and often sharper—business questions, and follow-up projects benefit from lessons learned in earlier cycles. At Twentynext, for example, insights gained in an R&D track on eye disease detection can directly shape the modeling approach for a later project on early tumor diagnosis.

CRISP-DM in generative AI projects

For modern generative AI implementations, CRISP-DM still works—but the way you fill in the phases changes. The data phase shifts toward prompt engineering and RAG architecture. Deployment now also includes hallucination monitoring, prompt version control, and governance. Twentynext adapts the methodology to fit the project type without changing the core principle: start with the business challenge, validate continuously with stakeholders, and document every iteration.

The migraine prediction project shows this clearly. The starting point was patient value: helping people gain more control over attacks through better prediction. From there, the team worked iteratively, with clinical feedback after every modeling sprint. That is CRISP-DM used the way it was intended.

What CRISP-DM adds in terms of transparency

For clients, this structured approach offers a practical advantage: at any point in the project, it is clear which phase the work is in and what deliverables should come with it. That avoids the all-too-common situation where, after three months, a client asks, “So what exactly has been delivered?” and the answer is a Jupyter notebook full of notes.

If you want to dive deeper into how CRISP-DM works in practice, Twentynext has a detailed explanation on its CRISP-DM page.

What you can do:

• In your next data science project, ask which phase the team is currently in and what deliverable belongs to it
• If the answer is vague, use CRISP-DM as the discussion framework: business understanding, data understanding, preparation, modeling, evaluation, deployment
• Check whether the process allows the team to move back to an earlier phase without bureaucratic delays; if not, the operating model is probably too rigid

From research paper to working system: what it looks like in practice

So what does an R&D week at Twentynext actually look like? It starts not with a technology choice, but with a question. Take the project focused on detecting and classifying tumor cells for digital pathology. The starting point was a clinical challenge: pathologists struggle to assess IHC images consistently because color variation between labs makes interpretation harder.

From paper to prototype

A data scientist reads a recent paper on adaptive color analysis in pathology images. The next step is not a fully worked-out project plan, but a fast prototype: does the underlying principle hold up on real image data? Within a few days, there is a first version that analyzes IHC images for tumor characteristics such as size, location, and cellular features.

That prototype is not meant for production. Its job is to answer two questions: does the technical assumption hold, and is the result clinically relevant enough to develop further? If the answer to both is yes, the team moves into deeper iteration through CRISP-DM.

From prototype to production code

The journey from prototype to production code takes longer than the first week, but the direction is set early. What follows is a structured walk through the CRISP-DM phases: data preparation using real clinical image sets, model evaluation with pathologists acting as domain experts, and ultimately deployment in an environment that meets Twentynext’s ISO-certified service and maintenance standards.

That final part matters. An AI model in production needs performance monitoring, data drift detection, and scheduled retraining. Without that infrastructure, even a strong model can become unreliable within six months—and no one notices until the damage is done.

You can see a similar pattern in the tumor cell detection project: the architecture is flexible enough to handle both IHC images and conventional biopsies, an expansion that started as a relatively small technical exploration during an earlier iteration.

What you can do:

• Treat a prototype and a production system as two different things: prototypes test assumptions, production systems must run reliably in a live environment
• For every AI project, ask who owns model performance monitoring after go-live
• Check whether there is a retraining plan; models trained in 2024 on data from 2022 will usually perform worse in 2026 than they did at launch

Industrial AI: R&D in manufacturing

R&D at Twentynext is not limited to healthcare. The company also works on manufacturing challenges such as CAD automation and lightweight construction. Here too, the starting point is not technology—it is a business problem.

CAD/CAM automation as a real-world example

Stairlifts are custom-built because every staircase has different dimensions, angles, and gradients. That makes the design process time-consuming, while customers often need short lead times—frequently because of medical urgency. Twentynext developed an AI-powered software module for AutoCAD that automates specific engineering design tasks based on precise staircase measurements. The result: engineers spend less time on repetitive design work and more time on complex exceptions.

Three techniques, one module

In the project with MasterShip Software, Twentynext combines three complementary techniques: machine learning for pattern recognition in existing designs, genetic algorithms for optimizing design variants, and a rule-based inference engine for applying engineering rules and compliance requirements. Bringing those three techniques together in a single module was not an accident. It was the result of an R&D track in which each technique was prototyped and evaluated separately before integration began.

For data engineers and AI specialists, this kind of work is exactly what makes a specialist firm in the Brainport region attractive: direct collaboration with industry partners, visible impact in a growing technology field, and no waiting years for internal approval.

What you can do:

• In AI projects for production or engineering environments, ask the scale question up front (how many designs per hour, how many variants in parallel) before choosing the architecture
• Always combine domain expertise (engineering rules), data expertise (machine learning), and software engineering (production-ready modules); if one of the three is missing, the project usually stalls
• Ask whether standard tools already in use—such as AutoCAD or ERP systems—can serve as the integration point; that lowers the adoption barrier considerably

Comparison: how different R&D models work in practice

Different ways data science firms integrate R&D:

Best practices checklist for Data Science, AI consulting, and Business Intelligence

Best Practices Checklist:

• Start with the business question, not the technology: An AI model without a clearly defined business challenge usually leads to a solution nobody ends up using.
• Use CRISP-DM as an iterative framework: Going back to an earlier phase is not failure—it is part of the method; document every iteration.
• Separate the prototype from the production system: Prototypes test assumptions; production systems require monitoring, drift detection, and retraining plans.
• Put model operations in place before go-live: ISO-certified service and maintenance processes, like those used by Twentynext, help prevent a good model from quietly becoming unreliable over time.
• Bring domain expertise into every phase: In medical AI, that means pathologists or clinicians in the evaluation phase; in industrial AI, it means engineers who understand production rules.
• Build governance in from the start: Generative AI implementations without an AI governance framework—including hallucination monitoring and audit logging—are a risk in production-ready environments.
• Measure scalability explicitly: Before choosing an architecture, define how many inferences per unit of time the system must support.
• Reserve R&D time alongside client work: Teams that operate only on a project-by-project basis usually fall behind on technology developments in their field.

What to avoid in R&D-driven AI projects

A few recurring patterns repeatedly show up as causes of delay or failure in practice.

Starting with the technology

The most common problem is this: a team starts with a technology choice (“we’re going to use a large language model”) instead of starting with a question (“how can we shorten the turnaround time for clinical reporting?”). It sounds like a subtle difference, but it shapes the entire project. Twentynext sees this pattern regularly in organizations that want to implement AI before they have clearly defined the business challenge. If you want to explore that topic further, see the article when is your organization really ready for an AI use case.

R&D without a feedback loop

R&D that stays entirely internal, without structured input from domain experts, tends to produce interesting prototypes that nobody uses. At Twentynext, the connection with university medical centers and industry partners is not window dressing—it is a functional part of the R&D cycle. Clinical feedback after every modeling sprint keeps the team from continuing down a path built on an assumption that stopped being valid three iterations ago.

Deployment without ongoing management

Once a model goes live, it needs to be maintained in production. Continuous monitoring and regular model tuning are often essential. And yet this is exactly the piece many firms leave out. Twentynext invests in ISO-certified service and maintenance processes because this is where most AI implementations are at their weakest.

What you can do:

• Before a project starts, confirm that there is a post-go-live management plan: who monitors model performance, who retrains it, and who escalates if performance degrades?
• Ask how domain expert feedback is organized: are clinicians, engineers, or other specialists structurally involved in the evaluation phase?
• Avoid starting with a technology choice; first define the business question in one sentence that a non-technical stakeholder can understand

Frequently asked questions

How quickly can Twentynext turn a new idea into an initial prototype?

In practice, the prototyping phase at Twentynext usually takes three to seven working days for an initial technical validation. That short cycle is intentional: the goal of a prototype is to test a technical assumption, not to build a production system. Only once the prototype shows that the approach works—and is clinically or industrially relevant—does the full CRISP-DM cycle begin, with all associated documentation and domain expert input.

What does working with university medical centers add to R&D projects?

University medical centers provide clinical domain expertise that data alone cannot replace. Without that expertise, it is impossible to determine whether a model is clinically valid, even if its technical performance looks strong. In projects such as the tumor cell detection platform and the eye disease detection system, the clinical feedback loop ensures the model is not only technically accurate, but also aligned with the realities of diagnostic practice. That is the difference between an interesting experiment and a useful system.

How does Twentynext support AI models after go-live?

ISO-certified service and maintenance processes are the backbone of Twentynext’s post-deployment approach. This includes model performance monitoring, drift detection (flagging when the data the model sees in production starts to differ from the training data), periodic retraining, and structured incident response. For healthcare organizations and manufacturing clients with strict continuity requirements, that is a meaningful point of difference compared with firms that only deliver project work and walk away after handover.

Why is CRISP-DM a good fit for both R&D and client projects?

The strength of CRISP-DM lies in its mix of structure and flexibility. The six phases provide clarity and transparency for clients, while the non-linear setup makes it possible to jump back to an earlier phase when new insight demands it. That flexibility is especially valuable in R&D: a discovery during modeling can fundamentally reshape the business question. In client projects, the same structure produces clear deliverables per phase and reusable lessons for future work.

How can data engineers and data scientists grow through R&D at Twentynext?

R&D accelerates growth at Twentynext because employees are involved early in external projects with meaningful technical challenges. Junior data scientists usually contribute to R&D projects alongside regular client work within three months. That speeds up technical development, but it also gives people context: someone who understands why a color analysis algorithm matters clinically will make better decisions faster than someone optimizing purely for model performance. You can read more about growth paths at a data and AI consultancy in the article on growing from consultant to tech lead.

Conclusion

Twentynext’s R&D approach is not a marketing story about innovation. It is a structural choice that shows up in how employees spend their time, how projects are designed, and how models are maintained after delivery. CRISP-DM guides every iteration, the company works closely with university medical centers and industry partners, and ISO-certified service and maintenance processes help ensure production models remain reliable after go-live.

For clients in North Brabant and beyond, that means working with a partner that can do more than build technically sound solutions. Twentynext also understands why a given technique matters in a clinical or industrial setting. And it always starts in the same place: what is the business problem? The technology comes after that.

Organizations that want to see how Twentynext puts this approach into practice can find an overview of projects, services, and methodology at twentynext.nl.