Fiducial lock · deterministic inference

A new model architecture for deterministic tasks…

Highest accuracy, precision and consistency for OCR, audio understanding, and structured data extraction.

Start building Docs

CAM 02 · PLATE A · LOCK CONF MEAN 0.99 SPECIMEN · NOT A LIVE DOCUMENT FAMILY NAME QUINN GIVEN NAME RIVER J. DATE OF BIRTH 1991-04-17 DOC NO. K-4418-Q InferAPI extraction overlay · bounds + confidence 0.99 lock 0.98

Works with any AI SDK

Official SDKs for TypeScript and Python, and a chat-completions API compatible with every framework out of the box.

Code sample

Schema in, fields out. Zod (or JSON Schema) describes the document; the model returns typed values with optional bounds and confidence.

extract.id.ts chat.completions · response_format
import { InferAPI, responseFormat } from "inferapi";
import { z } from "zod";

const client = new InferAPI({ apiKey: "<your-api-key>" });

const IDSchema = z.object({
  family_name: z.string().describe("Family name on the specimen"),
  given_name: z.string().describe("Given name on the specimen"),
  dob: z.string().describe("Date of birth"),
  doc_no: z.string().describe("Document number"),
});

const response = await client.chat.completions.create({
  messages: [{
    role: "user",
    content: [
      { type: "text", text: "Extract the details from this ID" },
      { type: "image_url", image_url: { url: "https://inferapi.quest/examples/specimen.jpg" } },
    ],
  }],
  response_format: responseFormat(z.toJSONSchema(IDSchema), "id_schema"),
});

Verifiable fields

Data you can put in a rule.

Every extracted field can carry a value, a confidence score, and bounding boxes — so a downstream system can accept, reject, or send a human the crop.

{
  "family_name": {
    "value": "QUINN",
    "confidence": 0.99,
    "bounds": {
      "top_left": { "x": 214, "y": 154 },
      "bottom_right": { "x": 346, "y": 180 }
    }
  }
}
FIELD family_name QUINN lock 0.99 · bounds attached

Print, speech, and image — in many scripts.

Extract and understand text, audio, and images across a wide set of languages. Same schema; local glyphs.

zh

托运单 4418 · 件重 12.4 kg · 目的地 大阪

hi

वेबिल 4418 · वजन 12.4 किग्रा · गंतव्य ओसाका

es

Guía 4418 · peso 12.4 kg · destino Osaka

ja

送り状 4418 · 重量 12.4 kg · 仕向地 大阪

Sandboxes and headless browse

Compute in a sandbox and fetch pages with a headless browser when the field lives behind a URL, not a scan.

Configurable guardrails

Policy packs for text and images: crime, sexual content, self-harm, PII, and election claims — on or off per key.

Architecture

Specialized extractors, then a transformer.

A hybrid stack: convolutional and domain nets for print, layout, and audio, with a transformer layer for language. Built so the same ID, form, or clip returns the same fields.

scan CNN / DNN print · audio transformer { lock }

Specs

PropertyInferAPI
ContextLong-context prompts and files
OutputStructured JSON or text
Input modalitiesText, images, audio, file, video
ReasoningAvailable when a field needs it
CachingIncluded
Observability & loggingComing soon
BillingUsage-based inference — request a key

Start building with InferAPI

Bring deterministic accuracy to your workflows. Get an API key — no credit card on the form.

Get API key Browse the SDK