The case for AI you actually own

Update, August 2026: half of this post is now wrong, and the half that's wrong is the half in the title. Marapone runs the engine and sells a subscription. The reasoning is at the bottom; the original argument is left below it, unedited.

Most AI products sold to businesses share a shape: your data goes up to their cloud, and a per-seat meter runs against you forever. For a lot of companies that trade is fine. For the ones we build for — construction firms sitting on proprietary drawings and bids, freight forwarders whose whole edge is their rate book — it's a non-starter.

So MaraponeAI started from a constraint instead of a feature list: nothing the customer feeds it ever leaves the building, and they pay for it once. Everything else had to be designed around that.

On-prem is a design decision, not a deployment note

When you commit to running on the customer's own hardware, a lot of comfortable assumptions disappear. You can't lean on a fleet of cloud GPUs. You can't ship a fix by redeploying a container nobody sees. The model has to be small enough to run on commodity machines and good enough that the size doesn't show.

That pushed us toward a quantised llama.cpp engine with a domain layer trained on real operational data, rather than a giant general model behind an API. It's less glamorous and far more useful: it runs offline, it's fast enough on hardware a firm already owns, and it's theirs.

You buy it once. It runs on your infrastructure. You get the source. Nothing leaves the building.

The privacy guarantee is the product

Once the engine lived entirely on-prem, the same guarantee carried into everything built on top of it — the construction suite reading Ontario Building Code drawings, the logistics suite auditing freight invoices. Same engine, same promise. GasperAI, the assistant inside both, never phones home because there's nowhere for it to phone.

The lesson we keep relearning: the hard part wasn't the model. It was having the discipline to say no to the cloud-and-subscription default that every other tool in the category reaches for.

What we got wrong (August 2026)

Two months after this went up, Marapone stopped selling the programs outright. They are a subscription now, from $49 a month, running on Marapone's infrastructure and opened in a browser. Two of the three claims in that quote above are gone: it does not run on your infrastructure, and you do not buy it once.

The failure was in the sentence I was proudest of — "they pay for it once" — and it wasn't a commercial failure. It was a product one. A building code gets amended. A statutory rate table gets re-dated. The model reading a drawing set is measurably better than it was six months ago. A copy sitting on a laptop from last spring gets none of that, and half of what we shipped started decaying the day it was delivered. The honest fix is to keep improving it, and a one-time price cannot pay for that. We were selling a snapshot and describing it as an asset.

The part that survived is the part that was actually load-bearing: the model is ours, fine-tuned by us, running on infrastructure we operate, with no commercial AI API anywhere in the path. Nothing a customer uploads trains anything, a workspace is scoped to one account, and firms whose contracts genuinely forbid documents leaving their environment get a deployment into their own tenant — negotiated up front, rather than promised on a marketing page to everyone and honoured for nobody.

What I'd write today is that "private" and "owned" were doing one job in this post and I was treating them as one word. Privacy is a property of where the model runs and who is in the path. Ownership is a property of a licence. Conflating them made for a better sentence and a worse argument, and it took shipping the thing to notice.