The plumbing under Little Bear. A small rack in your shop, running open-source models you own, on hardware you own, against data that never leaves the building.
Continuous ontology systems are token-eaters. Renting frontier APIs by the call turns a 5-person operation into a hostage to someone else’s pricing page. We install fixed-cost local stacks instead — quantized specialist models, deterministic tool use, your data on your floor.
MSP keeps the lights on. FDE earns the context. ONP holds the weights, the graph, and the writeback. Little Bear is the surface the operator touches. Take any one out and the loop stops compounding.
On-prem is not an aesthetic preference. It is the only configuration that satisfies all three constraints at once for an operations-heavy mid-market shop.
Job records, invoices, payroll signals, client correspondence — none of it leaves the building. No vendor logs your prompts. No third-party model trains on your operation.
Capex once, opex flat. The same agent loop that bankrupts you on per-call APIs runs all night for the cost of the electricity. Budget like grown-ups.
Models propose. MCP validates. SQL executes. The probabilistic surface is sandboxed inside a typed contract — it cannot send the wrong invoice to the wrong client because the schema will not let it.
The rack is not a side project for somebody else to babysit. The same engineers who watch your applications watch the rack — disk health, fan curves, model serving, weight versions, monthly operating review.
An ontology that mutates on every job, every form, every approval cannot be funded by a per-call API. The first month works. The third month bleeds. The sixth month is a budget crisis dressed up as a roadmap question.
For a 5-person back office, the rack pays for itself before the second quarter. After that, the line is flat. The API line never is.
The whole AI plumbing for an entire mid-market operation, stacked top to bottom. Open at every layer. Owned at every layer. Every layer is something we have racked before.
A 5-person back office does not need a server room. A small, well-chosen box in a locked closet runs the whole stack. We pick the route that fits the shop, rack it, harden it, and cover it under MSP — same engineers, same on-call, one throat to choke.
A single sealed box that sits on a shelf in the IT closet. Pulls less power than a space heater. No fan whine, no heat, no second mortgage on the electric bill. The right answer for most 5-to-50-person shops.
A standard tower with a serious GPU, the same shape as the box already under somebody's desk. Headroom for heavier workloads, hot-swappable parts, an upgrade path that doesn't require throwing the rack away.
Every install is sized to the actual workload — concurrent users, ontology size, agent loop frequency. We share the bill of materials with you. We do not share it on a public services page.
One enormous frontier model trying to do everything is the wrong shape for an ops backbone. We run a small ensemble of open-source specialists — each chosen for a single role, quantized for the rack, hot-swappable as the open ecosystem moves. The roles are stable. The weights underneath are ours to pick.
Writes the SQL, drafts the invoice, fills the form, calls the next tool. The model that turns an operator's intent into a typed action against your systems. Specialist-tuned for code and structured output.
Decomposes a fuzzy ask into a clean sequence of validated steps. Returns schema-compliant tool calls — the part the deterministic plumbing relies on. Best return on inference cost in the open ecosystem.
Resource-light work that runs constantly — log parsing, dispatch summaries, agent-to-agent messages, micro-classifications. Cheap to keep warm, fast to invoke, sized for always-on background loops.
Encodes the entire ontology into a local vector index for semantic recall. Re-embeds on a nightly cron without phoning home. The index is built, owned, and stored on the rack — never shipped to a third party.
We track the open-source frontier weekly and re-quantize new releases into the same role slots. The MCP contract above the models does not change when the weights underneath do — your stack improves quietly, in place.
The on-prem stack is not a separate engagement. It is part of the FDE embed. Our engineers rack the box the same week they start riding along. By the end of the embed, the rack is humming and Little Bear is mutating the graph.
You miss the frontier the way a working diesel truck misses a Formula-1 car. For ops work — schedule mutation, invoice compose, document parsing — quantized 14B–32B specialists clear the bar. We track new releases monthly, re-quantize, and hot-swap weights without changing the MCP contract above them.
Our MSP does. The same engineers who watch your applications watch the rack. Disk fails, fan stalls, vLLM crashes — alerted, ticketed, fixed. You are not running an in-house data-center team.
No. For long-tail tasks where Sonnet-class horsepower is genuinely needed, the gateway can fall through to a frontier API under explicit policy. The default is local. The exception is logged.
Three reasons: weights you can pin and audit, no surprise pricing email, no surprise license change. If a vendor pivots, your stack does not.
With prefix caching and continuous batching, yes. Static system prompts and stable ontology context get cached, so the per-mutation token cost approaches the cost of the diff. That is what makes background loops viable.
The MCP contract is the same regardless of what is behind it. Add a second card, a second box, or a second site — the tool layer does not change. The rack scales without rewriting Little Bear.
We will spec the rack for your shop, install it under the FDE embed, and hand it to MSP for ongoing coverage. Fixed cost. No vendor lock. Little Bear humming on hardware you own.
Spec a rack →