Two Problems One Solution: Local LLMs

Transparency statement: I currently work as a Defense Director for Altair Engineering and manage the Army Ecosystem (Primes and Direct). My analysis may naturally align with some of Altair's technologies/methods, but these are my opinions. My current role and experiences give me a ground view of implementing commercial technology to advance US Defense objectives. This weekly Newsletter is a?practical guide?to applying new technology in the?Defense industry.

Problem #1: Government agencies using commercially available LLMs— even those from US companies—face catastrophic security risks.

Problem #2: Defense primes share those risks but also have incentives to strengthen their IP and increase company valuation.

ONE SOLUTION: Gradually develop and deploy a private LLM trained on internal and historical data, turning proprietary knowledge into a market differentiator and allowing our defense industry to outpace competitors.

Building an LLM locally may sound daunting, but it is possible—maybe even essential.

Cutting Through the Noise: AI in Practical Terms

I've helped primes and government agencies deploy AI and related technologies for years. The field is flooded with buzzwords:

• AI, Machine Learning, Deep Learning, Large Language Models (LLMs), Generative AI, Agentic AI Data Analytics, Knowledge Graphs, Digital Twins . . .

Executives/Managers/Engineers don’t need definitions—they need real-world applications with a clear return on investment (ROI). Skip the semantics and focus on solutions that deliver value today.

We can leave the semantic argumentation to the Ph. D.s about whether or not what we deploy is "AI." Does it do what you want it to do? PERIOD.

The temptation around the excitement of applying an AI is to try and do it all at once. Ultimately, this results in classic paralysis by analysis and little practical application.

My recommendation is ALWAYS to try to reverse the paradigm. Deploy advanced analytics methods and focus on PRAGMATIC applications with real-world ROIs. If it ends up being "AI" that is a happy coincidence.

Data, Data Everywhere—But Where's the Intelligence?

Companies and governments have spent billions on R&D, creating vast amounts of data in reports, CAD files, testing results, and other formats. Much of this data remains locked in siloed systems. Or physically locked away in dusty vaults.

An internal LLM, powered by a structured semantic layer, can:

1. Extract value from historical data—turning past projects into actionable insights
2. Multiply the impact of existing IP—connecting decades of work into a cohesive knowledge base
3. Preserve institutional knowledge—ensuring expertise isn't lost to retirements or turnover

Digitization is not new, and it has been happening for years. The missing piece is a connected ontology* that enables AI-powered decision-making without duplicating sensitive data.

Ontology, in the context of data and AI, refers to a structured framework that defines the relationships between concepts, entities, and data within a domain. It provides a shared vocabulary and logical structure that enables machines to interpret," integrate, and reason over information effectively.

Why Not Just Use Commercial LLMs?

Public LLMs (like ChatGPT and DeepSeek) have demonstrated their practical capabilities, but deploying them internally to sensitive data is a different challenge. Security concerns make commercial models untenable for high-stakes industries.

I wrote an article about commercial LLMs acting as trojan horses a few weeks back.

Building an internal LLM may seem daunting, but focusing on practical, step-by-step implementation makes it achievable:

1. Structure your data – Deploy a tiger team to extract, transform, and load (ETL) your knowledge base.
2. Use a Knowledge Graph – Map relationships between historical projects, test results, and technical documentation.
3. Deploy a User Interface – Enable semantic search and intuitive access to enterprise knowledge.
4. Scale over time – Repeat this process across departments, gradually training an internal AI ecosystem.

When done right, an internal LLM doesn't just retrieve information—it connects the dots, enhances innovation, and accelerates product development.

A graduated approach toward graphing and deploying your data will multiply the value of existing IP and government-sensitive data. You don't need to abandon your existing data; in fact, you can leverage it to train your models and create market differentiation.

Example: Leveraging Generations of Engineering Data

Take an aerospace company that has designed and tested engines for decades. Their data is scattered across:

• Legacy CAD files
• Hand-drawn schematics
• Testing results in isolated reports
• Supplier specifications in PDFs

By systematically structuring this data and embedding it in a secure AI ecosystem, the company can:

1. Reduce design iteration time by referencing past engineering decisions
2. Accelerate new product development with AI-driven recommendations
3. Increase valuation by making IP a scalable, reusable asset

Over time, as more data is structured, patterns emerge, and the internal AI becomes a true force multiplier.

Internal AI = Enterprise Acceleration

Companies and Governments that fail to structure their data for AI integration in a PRACTICLE MANNER will fall behind. Those that do will:

1. Develop products faster
2. Reduce costs through optimized decision-making
3. Outpace competitors by unlocking the full potential of their existing IP

The technology is ready. The competitive advantage is real.

The next frontier of enterprise value isn't just using AI—it's building an AI-driven knowledge infrastructure that grows with the company.

Final thoughts

This is not easy, and it requires the expertise of dozens of people, institutional/cultural change, and a cold, challenging budget. But it is possible—maybe even essential.

But partners in this area are some of the best in the business: Sam Arnold ," Sam Chance , David Smalley

#AI #LLM #EnterpriseAI #DataSecurity #DigitalTransformation #IPProtection #KnowledgeGraph #Ontology #ArtificialIntelligence #MachineLearning #DeepLearning #GenerativeAI #SecureAI #DefenseTech #AerospaceEngineering #DigitalTwins #FutureOfWork #Innovation #BigData #TechLeadership #Altair #onlyforward