R&D Principles & Convictions

Bridging the gap between exploration and industrialization.

In industrial settings, research and development (R&D) must operate under a different set of rules than in academia. While academic research is measured by publications and novelty, industrial R&D is measured by shipped, usable product value.

Throughout my career, I have observed a recurring tragedy: brilliant β€œpotential innovations” representing months or years of intense intellectual effort that never leave the laboratory. This gap between discovery and delivery does not happen because the algorithms are wrong; it happens because of a lack of structural alignment and architectural foresight.

When a team operates in isolation, optimizing theoretical metrics without a clear path to the product, it leads to missed market opportunities and deep frustration for the researchers involved.

To bridge this gap, I follow and enforce four core operational principles.


1. End-to-End Integration Over Local Optimization

The Problem: R&D teams often spend months micro-optimizing an isolated algorithmic component (a β€œbrick”) before verifying how it interacts with the rest of the system. This premature optimization leads to mismatched APIs and wasted effort.

My approach is to build the entire end-to-end chain first, using a β€œnaive brick” in place of incomplete algorithms. A naive brick is a simple, existing, or basic solution in the codebase that can be implemented quickly (e.g., a simple heuristic or a standard library call).

By running this end-to-end flow from day one, we achieve two critical goals:

  • Interface & Data Contracts: We immediately define the necessary inputs and outputs required for each step. When considered as a whole, the data flow and boundary interfaces of a system are often very different than if you designed each component in isolation.
  • Process Justification: It forces us to prove if a proposed algorithmic step is actually needed and why, saving weeks of unnecessary development.

Once the complete pipeline is connected up to the final deliverable, we profile the system, identify the true bottlenecks, and iterate on optimizing them. This tracer-bullet architecture gives Product and Business teams early visibility into real-world performance and market readiness, while preventing engineers from solving problems that do not exist.


2. Structured De-Risking: POC, Prototype, and Production

The Problem: Without clear milestones, research projects easily drift. Teams either spend too much time building production-grade code for unfeasible ideas, or they test unintegrated scripts on users.

To prevent R&D teams from sinking time into unfeasible paths, I implement a strict three-phase gatekeeping workflow with clear exit criteria:

Phase 1: The Proof of Concept (POC)

  • The Core Question: Is this mathematically and technically feasible?
  • The Method: Built using separate, standalone programs, scripts, or off-the-shelf software tools. It is a quick demonstration of feasibility with no concern for code integration, optimization, or speed.
  • Exit Gate: Feasibility is proven.

Phase 2: The Prototype

  • The Core Question: Does this actually solve the client’s need?
  • The Method: An end-to-end software pipeline integrated into the product flow. It may contain bugs and is not yet optimized, but it serves as a functional vehicle for the Product team to test against real-world scenarios and gather user feedback.
  • Exit Gate: Product validation is secured. We never optimize a prototype that hasn’t passed this gate.

Phase 3: Production-Ready

  • The Core Question: Is it robust, fast, and maintainable?
  • The Method: The validated prototype is rewritten in the destination language (e.g., porting Python/PyTorch prototypes to optimized C++ or WebGL), adhering to strict coding principles, and subjected to rigorous equivalence testing against the prototype’s output to guarantee numerical correctness.
  • Exit Gate: The code is shipped.

3. Prioritizing Tool Maturity Over Hype

The Problem: Researchers and engineers are naturally drawn to the latest emerging models and frameworks. However, adopting cutting-edge technologies too early introduces risks of instability, bloated environments, and unneeded complexity.

I advocate for data realism and technological maturity. If training data is scarce or too difficult to collect in sufficient quantities, we do not force a complex machine learning model onto the problem. Instead, we prioritize non-training approaches, mathematical modeling, and classical geometry or heuristics.

Opting for mature, standard mathematical approaches brings distinct advantages:

  • Maintainability & Reusability: Established standards are easier to understand, document, and maintain. They can often be easily adapted to solve other adjacent problems in the codebase.
  • Cohesive Architecture: For clients and system design, using standard mathematical blocks allows us to link different product features to a common scientific core. This avoids building a β€œtentacular pile” of modern, disjointed technologies that do very similar things but require separate runtimes, environments, or heavy hardware dependencies.

4. Grounding Researcher Motivation in Real Impact

The Problem: If success in corporate R&D is only defined by academic novelty or theoretical perfection, researchers will retreat into academic isolation, losing touch with the company’s business goals.

Keeping highly talented, scientifically curious individuals motivated within a corporate environment requires a shift in how success is defined.

By embedding researchers directly in the process of building the end-to-end pipeline, they gain immediate exposure to the practical constraints and real-world problems of the field. When a researcher witnesses their code transitioning from a standalone script to a physical device, solving a client’s problem, and shipping to the market, their perspective undergoes a profound shift:

Solving a theoretical question that has never been answered before is deeply rewarding. But it is even more fulfilling to see that solution directly improve people’s lives.

Experiencing that impact first-hand is enlightening. It transforms R&D culture from one of theoretical isolation to one of shared, commercial engineering pride, proving that the highest form of innovation is the one that actually ships.

Arnaud Bletterer

Arnaud Bletterer

Research Scientist & Team Leader

Specializing in Computer Graphics, Image/Geometry Processing, and Machine Learning, with a passion for bridging the gap between mathematical exploration and robust industrial software. If you have thoughts on these topics, are facing similar R&D challenges, or simply want to exchange ideas, feel free to reach out.