Structured outputs,every time.
Particula-JSON turns documents and free text into schema-valid JSON. Constrained decoding means the schema you provide is the schema you get back, on every request. One small model on a single GPU inside your network: no vendor cloud, no per-token meter.
One job, done to spec.
Here is the spec.
Everything below is work the model handles on its own, with no wrapper, no retry loop, and no second call out to a larger model when it gets stuck.
Schema in, schema out
Decoding is constrained against the JSON Schema you send with the request. Malformed or extra-field output is structurally impossible, so parsers downstream never break.
Documents and free text
Invoices, emails, forms, and logs. A bundled pre-processor handles PDF layout and OCR text before the model sees it.
99.8% schema pass
Field-level accuracy on our held-out extraction suites. Our own measurement: the shipped harness reproduces it on your documents.
Fails loudly, never guesses
Fields the model cannot ground in the source come back null with a reason code, so silent fabrication does not reach your database.
Context, latency, hardware,
and what it takes to run.
The numbers a capacity plan actually needs. Sizing gets reviewed against your traffic before the pilot, and the figures go into the agreement in writing.
From first call to production
without a rewrite.
Three steps, and you keep the weights at the end of them. Nothing about the rollout depends on us staying in the loop afterwards.
- 01
Discovery call
30 minutes, no slides. Your task, your data shape, and the security constraints around it. If this model is the wrong fit, we say so on the call.
- 02
Pilot install
Containerized deployment to staging in your VPC or on-prem, on the same stack as Lumen and Notetaker. The eval harness runs on your data and the numbers go into the pilot agreement.
- 03
Production rollout
Runbook handoff, operator training, and 90 days on call. After that it is yours to operate, no vendor dependency.
Five others share
this runtime.
They run on the same serving stack, the same deployment path, and one API, so licensing a second model adds no operational surface.
“Bring a hundred real examples from your pipeline. We run the model on them live, the harness scores the output in front of you, and you decide with numbers instead of a sales deck.”
Sebastian Mondragon, Founder of Particula Tech