Under the hood
Most AI vendors hide the architecture and sell the magic. We do the opposite. This page walks through the real stack, layer by layer, in enough detail that your IT person can check our work. That is the point.
How we choose components
Every piece of the stack is chosen on three rules. It must run on infrastructure you or we control, never on a public AI service. It must be open and inspectable, so nothing depends on our say-so. And it must be proven in daily production, because we run this exact stack ourselves before any client does.
The model, your data, and the agents all live in one environment under your control. Privacy is a property of the design, not a promise in a policy document.
Open-weight models, open-source runtime, documented data routing. Your IT or compliance reviewer can verify every claim on this page independently.
Our own business runs on this stack: the same models, the same runtime, the same agents. Client deployments get architecture we operate daily, not architecture we read about.
The stack, layer by layer
Read it bottom to top, the way it is actually built: hardware first, then the model runtime, then the intelligence, then your knowledge, then the agents that do the work, then the door you walk through to use it.
You reach your stack through a private mesh network built on Tailscale, so it is never exposed to the open internet. Every connection is encrypted and device-authenticated. There is no public login page for anyone to find, probe, or brute-force.
WireGuard-based private networking · device-level authentication · no open ports
The working layer: a team of scheduled agents for inbox triage, research, drafting, pipeline, and reporting, each with a narrow, specific job. They run on a schedule, read from and write to your knowledge base, and report what they did. Agents are built fail-loud: when something breaks, you get a report, never silence.
Narrow scope per agent · scheduled runs · fail-loud error reporting
The layer everyone else skips, and the reason generic AI produces generic work. We turn your documents, records, and institutional knowledge into one structured, searchable private brain that the model and every agent read from. Your originals stay the system of record; the brain links to them, never replaces them.
Structured local knowledge vault · document pipeline with human review bins · originals preserved
The intelligence is an open-weight model, from families such as Qwen, Gemma, and Llama, selected and sized for your actual work rather than defaulting to the biggest thing that fits. Models run quantized with a 32,000 token context window as our standard configuration, so they can read real documents, not just snippets.
Open weights · quantized for efficiency · 32K context standard
Models are served through Ollama, the open-source runtime that has become the standard for running language models locally. It loads, serves, and swaps models on your hardware with nothing calling out. We run Ollama in production every day, which is why installing, tuning, and managing it for businesses is a core Airgapped service rather than a footnote.
Open source · local inference only · the runtime we operate daily
For on-premises installs, the standard build is a Mac Mini class machine on Apple silicon. Unified memory lets the model, the knowledge base, and the agents share one pool efficiently, and the machine itself is silent, small, and sips power, which matters when it lives in an office instead of a server room. Larger workloads step up to Mac Studio class hardware.
Apple silicon unified memory · office-friendly · sized to the work, not the hype
Honest numbers
Nobody publishes this honestly, so here it is from daily production experience. Sizes are approximate for quantized models and shift with quantization level and context length, but they are the right planning numbers.
| Model class | Memory footprint | What it is good at | Hardware fit |
|---|---|---|---|
| 7B to 9B | Roughly 5 to 7 GB | Everyday drafting, summarization, classification, and knowledge-base search. The fast lane for high-volume agent tasks. | Runs comfortably on a 16 GB Mac Mini |
| 13B to 14B | Roughly 9 to 11 GB | The workhorse class: stronger reasoning, better long-document work, reliable enough to anchor a multi-agent stack. | Wants 24 GB of unified memory |
| 27B to 35B | Roughly 15 to 20 GB | The hardest local work: nuanced drafting, complex synthesis, and the tasks where the extra capability is visible. | 32 GB and up, or Mac Studio class |
| 70B and up | 40 GB and beyond | Frontier-adjacent quality. Most small businesses do not need it, and we will tell you if you do. | Workstation or server class hardware |
The most common sizing mistake is buying for the biggest model instead of the actual work. A well-prompted smaller model with your knowledge base behind it routinely beats a bigger model working blind. We size during the Clarity Session, and hardware on owned builds is a separate at-cost line, so there is no incentive for us to oversell the box.
Two ways to run it
The full stack lives on a machine in your office. Nothing is sent to any cloud, not for processing, not for storage, not for the model. This is the configuration built for the most regulated work, where nothing can ever leave the building.
We provision, harden, and manage an isolated, single-tenant environment for your stack on infrastructure independently audited for HIPAA/HITECH and SOC 2 Type 2, including a firewall dedicated to your environment alone. Your work is processed only inside that environment, never by public AI services. For healthcare deployments, a business associate agreement is executed before any protected health information enters the environment.
The honest question
Ollama is free and installs in minutes, and if you want to tinker, you should. The install is the easy five percent. A business deployment is everything that comes after it.
What this architecture guarantees
Your information is never sent to shared, public-cloud AI services. The model runs inside your environment, on-premises or hosted.
Your data is never used to train any model, ours or anyone else's. Open-weight models are frozen; they read your data, they do not learn it.
Your stack runs in its own environment, on your hardware or in a single-tenant hosted environment with a dedicated firewall, never co-mingled with anyone else's.
Every deployment ships with a written map of exactly where your data lives and moves, so your IT or compliance reviewer evaluates the real architecture, not a marketing claim.
Technical questions
Firms like Airgapped do. We design, install, and manage the full private AI stack: Ollama-served open-weight language models on right-sized hardware, a private knowledge base built from your business data, and scheduled AI agents that do real work on top of it. The tools themselves are open source; the service is the sizing, the build, the data layer, and the ongoing management.
Yes, and for personal tinkering you should. Ollama installs in minutes. What a managed deployment adds is everything after the install: choosing the right model for your hardware and work, building the private knowledge base your business data lives in, wiring agents to actually use it, documenting the data routing for compliance review, and keeping the whole thing monitored, updated, and working. The install is the easy five percent.
Less than most people expect. A Mac Mini class machine with 16GB of unified memory runs models in the 7 to 9 billion parameter class comfortably, which covers everyday drafting, summarization, and search. 24GB handles the 13 to 14 billion parameter class that we use as the agent workhorse. Larger models in the 27 billion and up class want 32GB or more, or a Mac Studio. We size the hardware to the work during the Clarity Session, not the other way around.
For the biggest frontier models, no, and we will not pretend otherwise. For the work most businesses actually hand to AI, drafting from your own documents, summarizing, structured extraction, search across your own knowledge, a well-chosen open-weight model with the right context does the job, and it does it without your client data ever leaving your control. The honest comparison is not model versus model, it is what you can safely give each one. A private model can see everything; a public one should see nothing sensitive.
Yes. Ollama is the model runtime in our standard stack, serving open-weight model families such as Qwen, Gemma, and Llama. We run it in production daily, on our own infrastructure and our clients' deployments, which is exactly why we install and manage it rather than reselling someone else's black box.
Same stack, different location. On-premises, everything runs on hardware in your office that you own. Hosted, we provision and manage an isolated, single-tenant environment for you on infrastructure independently audited for HIPAA/HITECH and SOC 2 Type 2, with a dedicated firewall, and your work is processed only inside that environment, never by public AI services. For healthcare deployments, a business associate agreement is executed before any protected health information enters the environment.
Your AI. Your data. Your control.
Start with a Clarity Session. We map your workflows, size the stack to the actual work, and give you the plan in writing.
Book a Clarity SessionServing medical, legal, and financial practices across Middle Tennessee and beyond.