// use cases · document extraction

Turn documents into
structured data.

Read claims, invoices, contracts and forms, text, tables and scans, and return clean JSON. On EU infrastructure, fully auditable.

book_a_call

// how it works

From messy documents to clean fields.

Vision, extraction and classification from a single OpenAI-compatible endpoint, at volume, and only inside the EU.

step 01

Read any document

gemma4

Parse PDFs, scans and forms, text, tables and layout, including images, with a vision-capable model. No OCR pipeline to maintain.

step 02

Extract to your schema

deepseek-v4-flash

Pull fields straight into your own JSON schema with structured outputs, every value where you expect it, every time, ready to validate.

step 03

Classify & route

qwen3-embedding

Classify the document type and route it to the right workflow, with embeddings you can audit, confident matches first, edge cases flagged.

// drop-in

Change one line. Keep your pipeline.

Vision input and structured outputs work the OpenAI way. Change the base URL and key, point at gemma4, and get your schema back, privately.

read_the_docs

extract.py

from openai import OpenAI

client = OpenAI(
    api_key="sk-...",
    base_url="https://api.helmcode.com/v1",  # one line changes
)

# vision in, structured JSON out — straight into your schema
result = client.chat.completions.create(
    model="gemma4",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Extract the invoice fields."},
            {"type": "image_url", "image_url": {"url": invoice_png}},
        ],
    }],
    response_format={"type": "json_schema", "json_schema": invoice_schema},
)

// why helmcode

Extraction you can put in production.

The documents you process are full of PII and money. Closed APIs ask you to upload all of it, and log it.

Zero logs, by architecture.

The documents you extract, and the data inside them, are never stored, and never train a model.

Processed in the EU.

Invoices, claims and contracts stay on EU infrastructure, not on US hyperscalers subject to the Cloud Act. GDPR and AI Act native.

Vision + JSON, one API.

Read a scan and return validated JSON from a single OpenAI-compatible endpoint, no separate OCR vendor, no glue code.

No caps at volume.

Process millions of documents a month. Limits are RPM and concurrency per key, never total tokens, so a busy month isn't a surprise bill.

Open models, your schema.

DeepSeek V4-Flash, Qwen 3.6, Gemma 4. Your fields, your JSON, no proprietary extraction format to lock you in.

Drops into your pipeline.

OpenAI-compatible structured outputs and vision. Change the base URL and key; your extraction code keeps working.

In production across

Banking & fintech
Insurance
B2B SaaS
Public sector
Energy & utilities
Manufacturing
Legal

In production at

// extraction faq

Extraction, answered.

What operations and engineering teams ask before automating document workflows.

Can it read scanned PDFs and images, not just digital text?

Yes. gemma4 is vision-capable (and qwen3.6 is fully multimodal), so it reads scans, photos and forms, text, tables and layout, without a separate OCR pipeline.

Can I get the output in my own schema?

Yes. Use OpenAI-compatible structured outputs (response_format json_schema) so fields land in your exact JSON shape, ready to validate and store.

Do you store the documents I send?

No. Zero logs, documents and the data extracted from them are never persisted and never train a model. Extraction stops being a privacy liability.

How do I trust the output?

Structured outputs constrain the response to your schema, so fields are always present and typed. Validate per field and flag low-confidence cases for review.

Can it handle high volume?

Yes. There are no token caps, limits are RPM and concurrency per API key, so you can process millions of documents a month on predictable, flat pricing.

What about highly sensitive documents?

Run on a dedicated GPU or fully on-premise inside your own datacenter, the same API and code, with documents that never leave your network.

// get started

START BURNING TOKENS

Skip the AI infra work. Deploy your first private inference endpoint today.

Flat rate. EU data. OpenAI API compatible.