bq-mako series · information extraction

Entities & relations,
~75× faster.

Sound Prediction Mako Extractor reads documents and pulls out named entities and the relationships between them — at transformer-beating accuracy, on a single CPU.

No neural networks No attention No black box
~75×
Faster than a transformer, on CPUi
~700×
Smaller than a transformeri
CPU
No GPU. Runs on-premi
100%
Of outputs explainedi
vs. a fine-tuned transformer extractor (~300M params) · measured on CPU · hover any figure for detail
A new kind of model

Not deep learning. A new statistical method.

The Mako Extractor isn’t a smaller neural network — it’s a different kind of model entirely. It’s built on a new statistical technique born from a convergence of ideas in Bayesian inference and statistical physics: a method that is accurate and fast precisely because it is transparent, not in spite of it. You get a model you can read, audit, and run anywhere — not one you have to take on faith.

See it think

Every extraction comes with its reasons.

Most extractors hand you a black box: entities appear, relations appear, and you take it on faith. Mako shows the named evidence behind each decision — auditable by construction.

And this is not explainable-AI bolted onto a black box. Post-hoc methods like SHAP and LIME approximate a neural network from the outside — and can be unfaithful to what the model actually did. Mako has nothing to approximate: the explanation is the computation. Faithful by construction.

Input → extraction

A 54-year-old patient with type 2 diabetes was started on metformin 500 mg and lisinopril for hypertension after presenting with chest pain.

metformin 500 mg treats type 2 diabetes
lisinopril treats hypertension
patient presents_with chest pain
Why — the evidence

“lisinopril” → Drug

-pril suffix marks ACE inhibitors · matches a drug ontology · sits in a prescribing context after “started on.”

lisinopril treats hypertension

A Drug and a Condition joined by the cue “for” — the strongest learned signal for a treatment relation.

Every weight above is a named, inspectable feature. Nothing hidden in a tensor.
Input → extraction

Acme Corp acquired Nimbus Labs for $400 million, a deal led by Dana Reyes in San Francisco.

Acme Corp acquired Nimbus Labs
Dana Reyes led the deal
acquisition price $400 million
Why — the evidence

“Acme Corp” → Organization

Corp head-token · title-case shape · it is the subject of “acquired,” a corporate-action verb.

Acme Corp acquired Nimbus Labs

Two Organizations bridged by the verb “acquired” in subject–object order — a directed acquisition relation.

Same model, new domain — no retraining, just the type names you ask for.
Why teams choose Mako

Built for the places a transformer can’t go.

Fast & cheap

Runs in real time on a single CPU core — no GPU fleet, no per-token bill. Extract at the speed of your pipeline, not your budget.

🔎

Interpretable & auditable

Every entity and relation traces back to named features you can read. For healthcare, legal, and finance, “why” is not optional — and there is no neural net to explain away.

🔒

Private & deployable

A model small enough to run inside your environment, on-prem or air-gapped. Your documents never have to leave.

How it compares

Transformer accuracy, without the transformer.

Matches or exceeds fine-tuned transformer extractors on entities and relations — and on clinical text, it leads.

CapabilityMako ExtractorTypical transformer extractor
Speed (CPU)~75× fasterbaseline
Model size~700× smaller~300M params
Runs on CPU / on-premYesGPU-bound
Every prediction explainedYesNo (black box)
Entity accuracyMatches or exceedsbaseline
Clinical entity accuracyLeadsbaseline
Neural networks / attentionNoneRequired

Figures are current internal measurements; final published numbers to be confirmed.

Early access

Put it on your hardest documents.

Tell us what you need to extract. We’ll set up a demo on your data, or get you on the early-access list.

Prefer email? info@mederrata.com