CORE-MI — Computational R&D

01 · The Problem

Some problems don’t fit the standard toolbox

Real technical problems rarely arrive in the shape a standard package expects. Data is incomplete, noisy, unstructured or collected under conditions nobody designed for. Systems are coupled, nonlinear, or only partially observable. Requirements pull in different directions: fast enough to matter, interpretable enough to trust, adaptable without becoming impossible to maintain.

A forecasting model that works until the data drifts. An optimization heuristic that quietly ignores half the constraints. An AI system that performs well in the demo and unpredictably in production. It’s usually the same root cause: a generic method forced onto a problem it was never built for.

CORE‑MI exists for the problems on the other side of that gap, where the off-the-shelf solution isn’t good enough, and the difficulty is real rather than a matter of missing effort.

02 · The Approach

From problem to dependable system

“We don’t start with a technology — we start with the problem.”

01
Understand

Understand the actual problem, its constraints, the data available, and the outcome that actually matters.
02
Formulate

Translate the problem into a precise mathematical, computational or algorithmic formulation.
03
Explore

Investigate candidate approaches through mathematics, simulation, statistics, AI and experimentation.
04
Develop

Build the algorithm, model, software or computational system the exploration points to.
05
Validate

Test assumptions, performance, robustness and failure modes — not just the demo case.
06
Deliver

Produce a maintainable solution that integrates into the environment it was built for.

03 · Capabilities

A specialized toolbox, applied as one system

These are not independent service lines. They are disciplines CORE‑MI draws on together, in whatever combination a given problem actually requires. The technology follows the problem, not the other way around.

Mathematical Modeling

Formal representations of complex systems and phenomena as a foundation for analysis and computation.

Computational Modeling & Simulation

Numerical models and simulations for systems that are difficult or impossible to analyze in closed form.

AI & Machine Learning

Specialized ML and AI systems, built only when they’re genuinely the right tool for the problem at hand.

Data Science

Extracting structure, relationships and predictive signal from complex, high-dimensional datasets.

Optimization

Computational methods for constrained, nonlinear or otherwise difficult optimization problems.

Scientific Computing

Translating mathematical and scientific problems into computational systems built to run at scale.

Algorithm Development

Specialized algorithms designed from first principles when standard methods fall short.

Software Engineering

Turning algorithms, models and prototypes into software that can be maintained and extended.

Research & Development

Investigating problems where nobody yet knows what the right solution looks like.

04 · Selected Work

Representative problems

Most CORE‑MI engagements are confidential. The following describe the technical shape of each problem without disclosing proprietary detail.

A

Large-scale ballot recognition and classification

Problem: Recognize and classify handmarked ballots at scale, robust to scanning noise, ambiguous marks and real-world edge cases in election data.
Approach: A computer-vision and classification pipeline specifically tuned to the failure modes of ballot imagery, not a generic OCR stack.
Result: Clear gains in accuracy and processing speed for election data handling.
Learned: In a setting with legal and auditability requirements, the classifier’s uncertainty is as important as its accuracy.

B

An agentic coding assistant for embedded firmware

Problem: Firmware development for embedded systems and MEMS devices is slow and narrowly specialized. It’s bottlenecked by register-level detail, timing constraints and hardware-specific toolchains.
Approach: An agentic coding assistant scoped to the firmware domain, grounded in the target hardware’s real constraints: register maps, timing budgets, the actual toolchain.
Result: Faster firmware iteration, without handing final judgment on hardware correctness over to the model.
Learned: A coding agent is only as good as the domain knowledge and guardrails built around it. Once register-level and timing behavior are involved, general-purpose code generation stops being enough.

C

AI-assisted content generation for social media creators

Problem: Independent content creators need a steady volume of on-brand ideas and drafts without the cost of a full production team.
Approach: An AI-assisted content generation tool taken from concept to a working prototype, tuned to one creator’s voice and format instead of a one-size-fits-all template.
Result: A working prototype that cut the time from idea to publishable draft.
Learned: Generation quality on its own wasn’t the hard part. Holding a consistent voice across a high volume of output was.

D

Proof-of-concept work: live captioning/translation and ranking systems

Problem: Two earlier-stage engagements meant to test feasibility, not to build for production: real-time captioning and translation of live speech, and ranking and recommendation for personalized content.
Approach: Disciplined proof-of-concept builds — define the smallest experiment that would actually falsify feasibility, then build only enough system to run it.
Result: Clear, honest feasibility answers for both problems, with a concrete path to production scope where the answer was yes.
Learned: A proof of concept is a different deliverable from a prototype: the goal is a fast, defensible answer, not a demo that happens to look finished.

05 · Research & Space

Comfortable where scientific rigor is non‑negotiable

CORE‑MI’s founder, Miklós Kornyik, is Principal Investigator of the HUNOR IMU‑DRS — an inertial-sensor microgravity navigation — experiment he took from proposal through funding and integration to a successful run aboard the International Space Station, as part of the Ax‑4 mission (2025), in collaboration with the HUNOR Program and Hungarian astronaut Tibor Kapu.

Taking an experiment from proposal to flight and running a computational R&D company draw on the same discipline. It means defining the problem precisely, designing a valid test of it, and staying honest about what the results do and don’t show, across the years of work that actually sit between an idea and a result. That discipline, together with a completed PhD and published work on on various scientific topics, carries directly into CORE‑MI’s engagements, in aerospace or otherwise.

CORE‑MI is not a party to, and does not commercialize, the HUNOR IMU‑DRS experiment or its intellectual property. This section describes relevant experience of the company’s founder.

06 · About

An applied mathematician, not an AI vendor

Miklós Kornyik completed a PhD in Applied Mathematics at Eötvös Loránd University, with a thesis on random matrices and orthogonal polynomials. He also spent time as a teaching assistant at the University of California, Merced.

Over the past decade his work has spanned handwritten signature authentication, drug-discovery modeling, meteorological and energy forecasting, election-data classification, quantum random walks, and distributed-algorithm convergence theory. He’s done this for research institutes, including HUN-REN’s Alfréd Rényi Institute of Mathematics and Wigner Research Centre for Physics, as well as for industry partners. He founded CORE‑MI in 2023 to bring that range under one discipline: understanding a difficult problem before choosing a technology for it.

Outside CORE‑MI, he also performs and records as an indie guitarist under the name Mick Kornyik, with two albums released to date.

Have a problem that doesn’t fit the standard toolbox?

Tell us what you’re trying to solve.

Start a conversation

Computational solutions for problems that resist conventional approaches