AI where deterministic software stops working.
We don't use AI because it's fashionable. We use it only where ordinary software reaches its practical limits — and then we design and train models for that specific job, export to ONNX, and serve them on the CPUs you already own, which removes the hardware bill, the scheduling contention and the vendor dependency along with it.
Deciding whether two records describe the same real-world thing — the same apartment, building, vehicle or business. Names are spelled differently, addresses are formatted differently, photographs are taken years apart from different angles, and OCR and translation add their own errors. Rule-based matching either misses most of it or floods you with false positives, and the reason is arithmetic rather than effort: a set of 200,000 images contains roughly 20 billion possible pairs, so comparing everything against everything is not a slow approach, it is an impossible one. A model narrows that to a small set of plausible candidates; deterministic checks and a human decide which are real.
- Problem framing and feasibility review
- Dataset design and labelling strategy
- Model architecture and training
- ONNX export and CPU inference tuning
- Entity-resolution and matching pipelines
- Evaluation sets and accuracy reporting
- Deployment onto your existing cluster
- Human-in-the-loop review where it matters
A model and the pipeline around it — trained, exported for CPU inference, deployed on infrastructure you already run, with the evaluation set it was measured on so the accuracy number stays checkable.
Why not just use a GPU?
Because for a great many production tasks the model does not need one, and once it does not, the economics change: no GPU fleet to buy or rent, no scheduling contention, and inference runs on the cluster you already operate. Where a workload genuinely needs GPUs, we will say so.
Is this a wrapper around someone else’s model?
No. These are models designed and trained from scratch for the specific task — a lightweight CNN, in the case above — which is what makes CPU inference viable at all. A general-purpose foundation model would not fit the same budget.
Why not just use GPT, Claude or Gemini?
Because this is not a language problem. Deciding whether two photographs taken years apart show the same physical apartment is a narrow visual-similarity task, and a general model has never been shown the thing you are asking it about. Ours is trained on the specific domain, is measured on a held-out set from it, runs on your own hardware for a fixed cost per million images, and sends none of your data to a third party. Where a general model genuinely is the right tool, we will say so.
How do we know the accuracy figure is real?
You get the evaluation set it was measured on. The number is reproducible on your side, which is the only version of an accuracy claim worth publishing.