Engineering & Manufacturing

A Blueprint for the Self-Sufficient Engineering Company

Open-source language models now put the power to build software into the hands of the engineers who own the problem. The people who understand the physics, the machines and the process can build and improve their own tools, without tying that capability to any one model provider.

Open-source LLMsOpenFOAM & FEAManufacturing automationInstitutional knowledge
Blueprint-style illustration of the modern self-sufficient engineering company: an industrial robotic arm and open-source LLM processor linked to proprietary knowledge and a component surrounded by fluid streamlines.
Original blueprint-style illustration: manufacturing automation, open-source LLMs, and open engineering tools connected by company knowledge. The flow lines and component mesh are illustrative, not simulation results.
The central argument

Engineering and manufacturing companies no longer need to send every improvement to a software vendor or queue it behind a separate coding team. Open-source LLMs that can code, reason and use tools now let the engineers who understand a problem build internal applications, extend proven analysis tools and automate repetitive work themselves. The opportunity lies in the open-model ecosystem as a whole, not in any single brand.

The decisive advantage is not the model. It is the combination of a company’s own data, its accumulated engineering judgement, open tools and a far shorter path from idea to working software. Self-sufficiency means owning that path. It does not mean pretending that suppliers, software engineering or validation are no longer needed.

The knowledge-to-capability loop

The blueprint: your knowledge, your physics, your workflows.

Turn internal data and tribal knowledge into validated engineering and manufacturing capabilities the company owns outright.
  1. 01 / Capture

    Company knowledge

    Internal dataTests, drawings, production history, and failure records.

    Tribal knowledgeShop-floor experience, design rules, and hard-won exceptions.

  2. 02 / Build

    Engineer-led development

    Engineering & manufacturing physicsDomain expertise, physical models, and process constraints.

    Open-source LLMsCode, connect, automate, and extend proven engineering tools.

  3. 03 / Validate

    Measure, test & approve

    Physical evidenceExperiments, conservation checks, and representative operating conditions.

    Engineering reviewSoftware tests, safety checks, traceability, and human approval.

  4. 04 / Own

    Company-owned workflows

    Analysis & simulationFluid flow, heat transfer, stress, vibration, and other physics.

    Manufacturing improvementInspection, monitoring, automation, and process optimization.

    Retained in-houseCode, tests, documentation, and the ability to improve them.

Feedback to Step 01: measure, capture lessons, improve.Validated results, production outcomes, and reviewed lessons flow back into internal data and tribal knowledge for the next cycle.

Engineers own the assumptions and approvals. Open-source LLMs assist implementation; physical evidence establishes whether the result is fit for purpose.

In this article
  1. The blueprint: from knowledge to owned workflows
  2. The bottleneck is moving
  3. Engineers become tool builders
  4. Open models, frontier performance
  5. Better analysis without starting over
  6. OpenFOAM: serious physics you can own
  7. From the analysis desk to the factory floor
  8. The hundred-year advantage
  9. Fewer licences, not less responsibility
  10. Building the capability
  11. References & related reading

1. The bottleneck is moving

In most engineering organisations, ideas are not scarce. What is scarce is the capacity to implement them. The engineer already knows what needs fixing: the thermal report that takes half a day to assemble, the stress workflow that needs the same manual edits for every variant, the production problem whose data sits in three systems that do not talk to each other.

Until recently, closing that gap meant a software request, a budget discussion, a consultant or a place in someone else’s backlog, and many useful tools were simply never built. LLM-assisted coding changes the economics of that last mile. The engineer describes the workflow, supplies representative inputs and expected results, and refines a working version. An improvement that sat on a roadmap for years can sometimes become a usable prototype in days. That is a statement about removing queues and friction, not a claim that decades of numerical-methods work or product qualification have vanished; a file converter and a certified structural-analysis system are very different things.

It also connects to an earlier argument about the flaw in calling engineering tools “outdated”. A mature in-house code can embody decades of validated practice. The opportunity is usually to make it easier to use, automate, test and maintain, not to throw it away because its interface looks old.

The company already understands the problem. What has changed is how directly that understanding becomes working software.

2. Engineers become tool builders

Domain engineers do not have to become full-time developers to build useful tools. A mechanical engineer can build a fixture calculator, a test-data processor or a parameterised analysis workflow; an aerospace engineer can direct a load-case generator or a simulation-comparison tool from engineering requirements rather than a blank editor.

The division of labour is clear. The engineer defines the physics, the assumptions, the operating envelope and what counts as correct. The LLM helps write the interfaces, scripts, tests and documentation. Software specialists review architecture, security and maintainability wherever the stakes justify it. That reduces dependence on coding teams for every small change, but it does not make software literacy optional: engineers still need enough of it to read a change, understand its dependencies, recognise a failed test and know when to ask for a review.

The LLM-directed OpenFOAM study, the Turbulence Realm analysis tools and the Earth Intelligence project each show part of this shorter loop. They demonstrate what can be built; they are not evidence that every generated tool is ready for production.

3. Open models, frontier performance

The leading open-source and open-weight LLMs now compete at frontier level on selected coding, reasoning and tool-use benchmarks. Openness no longer means settling for a weak assistant, and that makes this strategy possible: strong implementation support combined with control over deployment, data and choice of model.

Open-weight LLMs illustrate the shift; they are not the strategy. One open-weight LLM releases the full weights of a multimodal, agentic model with a million-token context under a licence that is permissive for ordinary use, with conditions aimed mainly at very large service providers. Another LLM family takes a similar path: its flagship targets complex software engineering and long agent tasks, while a smaller, multimodal LLM variant is released under the plain MIT licence. Their published results show open models competing with leading proprietary systems, though performance varies by task and developer-reported benchmarks are not proof of reliability in every engineering workflow. [1] [2]

Proprietary assistants remain useful, but an independence strategy cannot rest on one hosted service. With suitable open models, a company can run approved inference inside its own environment, choose its architecture and swap models as capability improves, provided its workflows, evaluation cases and knowledge stay portable. Two cautions apply. “Open source” often really means open weights, without the full training data or pipeline, so the exact licence matters. And self-hosting a large model has real costs in memory, accelerators and operations; a smaller approved model can handle routine scripting while a larger one takes the difficult work.

The best way to choose is a small internal benchmark of the company’s own work: a test-data parser with known outputs, a CFD case-setup task, a refactor of an existing utility, and deliberately broken inputs that must be caught rather than silently accepted. Run every candidate model against it on each upgrade, record pass rate, cost, speed and failure modes, and protect the suite like any other proprietary asset. An afternoon of internal evidence is worth more than any public leaderboard.

Local weights are not the whole privacy boundary

Prompts, retrieved documents, logs, embeddings, backups and an agent’s tool calls all need controls. Restrict network access and permissions, protect export-controlled and customer data, and keep sensitive work off unapproved external services. A locally hosted model whose agent uploads files is not a private system.

4. Better analysis without starting over

The fastest route to value runs around an existing solver, not through a new one. LLMs can help build repeatable inputs, expose useful controls, connect datasets and standardise results, and they can assist with changes to source code too, though a new boundary condition or material model demands far more scrutiny than a report generator. The pattern holds across disciplines. In fluid flow, it means case templates, geometry sweeps, meshing pipelines and pressure-loss reports, checked for conservation, mesh sensitivity, turbulence assumptions and agreement with measurements. In heat transfer, thermal networks and cooling studies checked against energy balances and sensor uncertainty. In stress analysis, automated load cases and mesh studies checked for units, constraints, reaction balance and allowables. In vibration and dynamics, modal and transient workflows checked against damping assumptions and measured modes. And for legacy tools, reviewed wrappers, regression tests and a usable front end that preserve validated behaviour.

The building blocks already exist: OpenFOAM function objects for runtime post-processing, ParaView’s Python scripting, CalculiX for linear and nonlinear structural, dynamic and thermal problems, EDF’s industrially qualified Code_Aster for mechanics, dynamics and fatigue, and Elmer for multiphysics problems such as acoustics and electromagnetics. The LLM’s role is to connect and customise these proven capabilities, not to invent numerical physics from scratch. [3] [4] [5] [12] [13]

The distinction to keep in mind is between a better workflow and a more accurate model of reality. An automated mesh sweep improves consistency; a prettier contour plot proves nothing about accuracy, which is the lesson of the simulation-on-simulation trap. The same ownership can extend a level higher, too: open frameworks such as NASA’s OpenMDAO couple aerodynamics, structures, thermal and cost models into multidisciplinary design studies, with an LLM helping to write the wrappers while the couplings and physics remain the engineer’s responsibility. [14]

5. OpenFOAM: serious physics you can own

OpenFOAM is not a toy for simple demonstration flows. The OpenFOAM Foundation distributes a mature, free CFD platform, developed and maintained by CFD Direct, covering fluid motion, heat transfer, thermodynamics and chemistry, and its GPLv3 licence lets companies inspect and extend the source. In-house modifications need not be published just because they exist, though distributing GPL software brings obligations worth reviewing. [9]

Its documented capabilities reach well beyond single-phase aerodynamics: turbulence models from Reynolds-averaged closures to large-eddy simulation; multiphase flows, from sharp interfaces between immiscible fluids to dispersed phases exchanging heat and mass; particle-laden flows such as sprays, fluidised beds and filtration; conjugate heat transfer, natural convection, reacting flows and combustion; and rotating and moving geometry for mixers and impellers. [10] That gives an internal team a serious numerical foundation. The LLM can automate case preparation, parameter studies, post-processing and reviewed extensions, but model choice, resolution and validation still decide whether a prediction is fit for purpose.

Open source does not mean unsupported. CFD Direct offers optional professional help with modelling, meshing, numerics, parallel computing, automation and customisation, and industrial users can help fund and steer development through the Foundation’s Process Engineering Consortium. The company owns its workflows and buys expertise where it adds value. [11] [10]

Use documentation for the right distribution

The Foundation’s releases at openfoam.org and OpenCFD’s at openfoam.com are distinct distributions. Pin one distribution and release for an internal workflow, and give the coding assistant that version’s documentation rather than mixing solver names and syntax.

A workflow a company can own

Picture a team that repeatedly assesses pressure drop and cooling across a family of industrial components. (The example is illustrative, not a reported deployment.) It begins with a trusted case: a baseline with its units, geometry conventions, properties, operating conditions and test evidence, on a pinned OpenFOAM version. An LLM then helps write scripts that generate controlled variants from approved templates, keeping every run traceable to a revision. Bad runs are rejected explicitly, on mesh quality, exit status, conservation and the stability of the quantities that matter, because low residuals alone prove nothing. The repetitive analysis is automated with function objects and ParaView, with failed and inconclusive runs recorded rather than dropped. Results are compared with analytical cases and calibrated measurements, kept separate from any calibration data. And the whole thing is packaged with a simple interface, versioned environments, regression tests, a named owner and an approval path for any change to the physics.

Once reviewed, most of that pipeline should run as ordinary, deterministic software. Use the model to build the automation; do not add a model dependency where a tested script will do. And if a new correlation or material law is needed, the LLM can help implement and test it, but the derivation, verification and experimental validation remain engineering work.

6. From the analysis desk to the factory floor

The same approach reaches beyond simulation. Engineers who understand production can build useful automation alongside controls, quality, maintenance and IT colleagues, with tools such as Node-RED for event-driven flows and ROS-Industrial for manufacturing robotics, perception, calibration and motion planning. [6] [7] The possibilities include condition monitoring built on approved vibration, current, pressure and temperature data; inspection tools whose false accepts and rejects are measured on real parts; test benches that capture calibration data and produce traceable reports; process and energy studies tested in controlled trials rather than inferred from correlation; scheduling, bottleneck and traceability dashboards that begin with read-only access; and offline robot path planning and simulated work cells qualified on the real hardware.

MQTT and OPC UA offer integration paths where equipment supports them, and LLMs can write the adapters, but engineers must define what every signal means, its timing and units, who may access it and what happens when a connection drops. A convincing dashboard built on misread signals is worse than none.

Keep the safety boundary explicit

An LLM is not a safety controller or a deterministic real-time system. Emergency stops, interlocks and safe motion belong in properly engineered control systems. Start with monitoring and offline recommendations, and use simulation, staged commissioning, rollback and authorised review before any production change.

The posts on small-board sensing and automation and LLM-assisted home automation show how accessible the software layer has become. They are not industrial safety precedents; plants add reliability, cybersecurity, environmental and regulatory demands of their own.

7. The hundred-year advantage is knowledge, not data volume

For a manufacturer with more than a century of history, the archive may hold drawings, test books, process deviations, rejected parts, field failures, repairs and the reasons behind its design rules. Not every company has such a past, and a hundred years of operation does not mean a hundred years of clean digital data. But that accumulated knowledge can be extraordinarily hard for any outsider to reproduce.

Then there is tribal knowledge: the machinist who knows which fixture distorts a thin wall, the test engineer who recognises a sensor artefact, the metallurgist who knows why an apparently acceptable heat treatment fails for one section thickness. Much of a company’s competitive edge lives in exactly these exceptions. A general commercial tool must serve many customers; an internal one can encode a single company’s machines, materials, failure modes and accepted practice in far greater depth, and that is an advantage to demonstrate, not an automatic superiority.

The practical first step is not training a model from scratch. It is a curated, permission-aware knowledge system that retrieves approved documents and records with citations and connects them to deterministic calculators. That means preserving context, linking each record to part revision, machine, material batch, calibration and approval; capturing the reasoning behind rules from experienced staff, with their consent; separating measured evidence from habit; evaluating on independent data; and enforcing permissions so that restricted programmes and controlled data are never exposed just because they were indexed. NIST’s industrial AI work makes the same point, combining physics, data and human observation with careful attention to provenance and domain-specific evaluation. [8]

Validated simulation archives belong in the picture too. Once a reviewed CFD or FEA workflow produces trustworthy results, its runs can train surrogate and reduced-order models that explore design options in seconds, provided their training range is documented, their error is measured on held-out cases and they flag when they are extrapolating. Physics-informed models go a step further, building the governing equations into training and fitting what remains to the experimental and test data already sitting in company archives. They need far less data than purely statistical models and can capture real effects a chosen simulation model misses, as the earlier exploration of missing physics showed. [15] The acceptance rules are unchanged: split the measurements before training, test on conditions the model has never seen, quantify the error, and keep checking against new data, because, as the experimental-testing post-mortem showed, a plausible pipeline can produce confident, wrong answers.

8. Fewer licences, not less responsibility

LLM-assisted development makes it far more feasible to move selected, repetitive work onto open-source or in-house tools, and so to rely less on expensive commercial seats and modules. That is a credible path, not evidence that the industry has abandoned commercial software. Commercial codes still earn their place through specialist physics, support, interoperability, qualification history and customer acceptance. Keep them where they add value, automate them through supported interfaces, and replace narrowly defined workflows only when the alternative meets the same acceptance criteria.

The right comparison is total cost of ownership: licences and external development on one side; compute, integration, validation, training, maintenance, security and downtime risk on the other. Self-sufficiency does not mean using nothing built elsewhere. It means being able to inspect, adapt, maintain and migrate the parts that matter, while owning the company’s data, workflows, acceptance tests and accumulated improvements.

Buy expertise where it adds value. Do not rent back your own engineering knowledge because nobody had time to build the interface.

9. Building a capability, not a collection of demos

The sensible place to start is one repeated, measurable task with a known answer: a post-processing report, a test-data import, a read-only quality dashboard. Pair a domain engineer with reviewers who understand both the software and the plant systems it touches, rather than handing an unrestricted agent the factory network.

Measure the current effort, errors and turnaround, and set the criteria for correctness before any code is written. Keep the first tool tightly bounded, with approved data, limited permissions, a sandbox and explicit behaviour when things fail. Build it against known-good cases, edge cases and deliberately bad inputs, and review the code and its dependencies rather than the demonstration. Run it in parallel with the accepted method and investigate every discrepancy instead of tuning it away. Name the people who own it, keep its history and rehearse recovery. Then share what works across teams, and measure the time saved after review and maintenance rather than the amount of code generated.

Faster implementation makes independent checks more valuable, not less; as the limits of LLM-driven experimental testing showed, a fluent explanation can hide a bad measurement, a unit error or an unjustified assumption.

What excites me is not a future in which engineering companies no longer need software expertise. It is one in which the people who understand the product and the process can act on that knowledge directly, with software specialists supporting a far larger community of engineer-builders. The lasting asset is not the prompt or the model subscription. It is the growing library of validated workflows, usable company knowledge and in-house skill. That is how manufacturing companies become more self-sufficient: one owned, tested improvement at a time.


References & related reading

Primary sources for the model, software, and industrial-AI capabilities discussed above. The proposed workflows and business implications are engineering analysis, not reported results from these organizations.

  1. Open-weight LLM model card (1)Open-weight multimodal agentic model: 1M-token context, long-horizon coding and reasoning, and its own model licence. Review the linked license for the exact terms: conditions apply to large Model-as-a-Service providers and high-reach commercial products, not ordinary internal use.
  2. Open-weight LLM model card (2)Flagship LLM for complex software engineering and long-horizon agent tasks, with a 1M-token context under its own licence. The efficient, natively multimodal variant is released under the MIT license.
  3. OpenFOAM documentation: Function objectsRuntime and post-processing operations, batch workflows, and reduced manual interaction. This reference is for the OpenCFD v2606 documentation; match instructions to the installed distribution.
  4. ParaView: Batch Python ScriptingPython automation, trace recording, and the pvpython and pvbatch execution environments.
  5. CalculiX: A Free Software Three-Dimensional Structural Finite Element ProgramLinear and nonlinear FEA, static, dynamic and thermal capabilities, and source licensing.
  6. OpenJS Foundation & contributors: About Node-REDFlow-based programming, reusable integrations, and edge deployment.
  7. ROS-Industrial: Project descriptionManufacturing robotics, calibration and planning tools, standardized interfaces, and integration with industrial controllers.
  8. NIST: Industrial Artificial Intelligence Management and MetrologyPhysics, human observations, heterogeneous data, provenance, and domain-centered evaluation of industrial AI.
  9. The OpenFOAM Foundation: OpenFOAMThe Foundation distribution, its engineering scope, and CFD Direct's development and maintenance role. See also the Foundation's free software licence explanation for in-house customization and distribution obligations.
  10. The OpenFOAM Foundation: Chemical and Process EngineeringDocumented multiphase, turbulence, particle, heat-transfer, reaction, meshing, and mesh-motion capabilities, plus the Process Engineering Consortium's support for development and testing.
  11. CFD Direct: CFD SupportOptional professional support for complex CFD practice, physical models, meshing, numerics, parallelization, post-processing automation, and code customization.
  12. EDF: Code_AsterOpen-source finite-element solver for mechanics, thermal analysis, and dynamics, with documented vibration-fatigue and damage operators and an industrial qualification process at EDF.
  13. CSC: Elmer FEMOpen-source multiphysics FEM covering fluid dynamics, structural mechanics, heat transfer, acoustics, and electromagnetics, with demonstrated parallel scalability.
  14. NASA Glenn Research Center: OpenMDAOOpen-source Python framework for systems analysis and multidisciplinary optimization: decomposed coupled models, analytic derivatives, and gradient-based or gradient-free optimization across large design spaces.
  15. Raissi, Perdikaris & Karniadakis: Physics-informed neural networksThe foundational Journal of Computational Physics (2019) framework for solving forward and inverse problems by embedding nonlinear PDEs in deep-learning training, including learning from scattered experimental measurements.

Continue with the OpenFOAM demonstration, why mature engineering tools should not be dismissed, or where the LLM still needs a human.