Practical AI
Practical AI, on infrastructure you control.
Most AI projects start by moving your operating data into someone else's platform. We start with one workflow and build it on infrastructure you control — private models, answers that cite their sources, and results checked against your own data before anyone relies on them. Already on Microsoft 365 or Azure? We'll tell you when your own tenant is the right place to run it.
Engineering in practice
Cemetery Brain: self-hosted knowledge infrastructure
Internal case study. We moved a stalled document-indexing pipeline onto a GPU server we operate, restored unattended publishing, and measured about 16× the embedding throughput of the previous CPU-only setup (34.0 vs 2.1 passages per second in a 512-passage benchmark). The catalog reached 143,321 searchable passages on September 2, 2026. Answer-quality evaluation is the next phase.
Read the case studyWhy private
Your maintenance history, customer records, financials, and internal documents are the assets that make AI useful — and the ones you least want leaving your control. We design AI systems that run on hardware you or we operate directly. Your data is never used to train third-party models, there are no per-seat platform fees, and you own the code, data, and documentation — so a vendor changing its terms doesn't strand your system.
What we build
Private language model deployment
Open models running on owned or leased GPU infrastructure, sized to your workload and your budget, with your data staying inside your boundary.
Document and records intelligence
Search and question-answering over contracts, manuals, correspondence, and decades of records — answers that cite their sources, so staff can verify instead of trusting.
Forecasting and exception detection
Models measured against your own history before anyone relies on them. If a simple method wins, we ship the simple method and tell you why.
AI-assisted engineering and automation
Software and automation pipelines that use AI for the work and keep a human approval gate on anything that touches money, records, or production systems.
How an engagement runs
Start with one workflow. Build something useful. Keep it working.
Less manual reporting. Easier document retrieval. Faster handling of routine inquiries. Those are the outcomes; the models and the hardware are how we get there. Every engagement follows the same three stages, and each stage stands on its own — you can stop after any of them with something you can use. We use the same approach on our own operations first — the Cemetery Brain pipeline below is one example.
Stage 1
AI Opportunity Audit
Find the workflow worth doing.
A fixed-scope review of your operation: which manual reporting, document-handling, and inquiry-handling work is actually worth automating, what data you already have to do it with, what it would cost to run, and what to leave alone. You get a written plan you can act on with us or without us.
You get: A ranked, costed plan
Stage 2
Automation Build Sprint
Build one workflow, end to end.
One workflow from the audit, built and put into daily use: the report that assembles itself, the records your staff can question and get cited answers from, the routine inquiries handled before a person needs to step in. Measured against how the work is done today, with a human approval gate wherever money, records, or production systems are touched.
You get: A working system your team uses
Stage 3
Managed AI Operations
Keep it working.
Once it works, we keep it working: monitoring, model and data updates, accuracy checked against real outcomes, and cost control on the infrastructure it runs on. Scoped to what you actually run, on hardware you or we operate, with no platform lock-in.
You get: Reliability you do not have to staff
Scope, fees, and timelines are set in a written proposal once we have seen your operation — the first conversation is how that starts.
Talk to Andrew about a first workflowHow we work
We measure before we recommend
Forecasting models are backtested against your own history; search and question-answering systems are tested against a set of your real questions before launch. When a simpler method wins, we ship the simpler method and show you the comparison.
Humans stay in the loop where it counts
AI drafts, analyzes, and flags. People approve anything that moves money, alters records, or changes production. That boundary is designed in from the start.
Grounded answers, not confident guesses
Retrieval systems cite sources and are built to say "not in the corpus" rather than invent an answer. A system that admits uncertainty is more useful than one that never does.
Engineering economics first
High-density compute is an electrical problem before a software one. Power, cooling, and total operating cost get modeled before capital is committed.
Everything described here is work we design, build, and operate ourselves. Each engagement is scoped in writing.
Wondering if AI is worth it for your operation?
That's the right question — and sometimes the honest answer is "not yet." Start with a direct conversation with the engineer who will tell you either way.