Sparrow GitHub Contact Try Sparrow →
Katana ML

Structured Data Extraction
& Agentic AI
Structured Data
& Agentic AI

Local inferencePrivacy-firstProduction-grade

Sparrow is an open-source AI platform by Katana ML for structured document data extraction, instruction calling, and agentic workflows — combining ML, LLM and Vision LLM with custom parsing, validation and query logic, all on your own infrastructure.

sparrow.katanaml.io
Sparrow document extraction interface — upload, extraction schema, and validated JSON response

Upload a document. Define a schema. Get structured JSON.

Try it live
The platform

More than extraction. A document-AI stack.

Three composable pipelines — from structured extraction to instruction calling and autonomous, multi-step agent workflows.

01 — Vision LLM

Sparrow Parse

Schema-validated JSON from invoices, statements, forms and tables — Vision LLMs backed by custom table parsing and validation logic.

02 — Text LLM

Sparrow Instructor

Instruction calling beyond extraction — text processing, analysis, field validation and decision-making.

03 — Orchestration

Sparrow Agents

Chain classification, extraction and validation into monitored, multi-step agentic workflows.

Why Sparrow

Built for control, not convenience.

01

Local Inference Only

Documents never leave your infrastructure. No cloud API calls, no third-party data routing. Full control.

02

Pluggable Architecture

Mix Vision LLM, OCR, and agent pipelines. Runs on Apple Silicon, NVIDIA GPU, or CPU via vLLM or MLX.

03

Enterprise Ready

In production with enterprise clients. Rate limiting, usage analytics, API-first design, commercial licensing available.

How it works

Three steps to structured data.

01 / Upload

Drop a document

Images (PNG, JPG) or multi-page PDFs. The file is removed after inference completes.

02 / Define schema

Describe the fields

Provide a JSON schema listing the fields you want to extract and their expected types.

[ { "instrument_name": "str", "valuation": "int" } ]
03 / Get JSON

Validated output

Sparrow returns structured, schema-matched JSON. Ready to drop straight into your pipeline.

{ "valuation": 213030, "valid": "true" }
Extraction hints

Rules that teach the model how to read.

Hints are plain-language rules that steer extraction on complex documents — focusing attention, disambiguating look-alike fields, normalizing formats, and even deriving values that never appear on the page.

  • Steer attention to footers, totals and fine print.
  • Disambiguate structurally similar fields — supplier vs. recipient VAT.
  • Normalize date and number formats, and resolve priority for ambiguous fields.
  • Derive missing fields — e.g. a risk flag on a bond position that isn't printed.
# rules that steer how the model extracts hints = { "supplier_vat": "VAT beside 'Supplier', not 'Bill to'", "issue_date": "normalize to YYYY-MM-DD", "risk_flag": "derive: 'high' if maturity > 10y" }
Agentic AI

Agents that reason over documents.

Sparrow chains steps into autonomous workflows — classifying, extracting, validating and deciding, with every run monitored visually.

01 · CLASSIFY

Identify

Detect the document type and route it to the right pipeline.

02 · EXTRACT

Parse

Vision LLM reads the fields; custom logic reconstructs tables and line items.

03 · VALIDATE

Check

Validate values against the schema; flag and retry on mismatch.

04 · DECIDE

Act

Apply business logic — like tax code prediction — and hand off clean results.

Custom processing logic handles complex layouts like bank statements — the multi-row tables pure LLM output can't — and every workflow run is monitored visually.
Coverage

Works across document types

From simple receipts to complex multi-section financial documents.

Invoices
Bank statements
Receipts
Financial tables
Tax forms
Custom business documents
sparrow.katanaml.io/dashboard
Sparrow dashboard showing inference volume, success rate, latency and model distribution
In production
Sparrow runs in production with enterprise clients, handling invoice extraction, field validation and tax account code prediction — from Vision-LLM parsing to custom bank-statement processing and agent workflows, on local infrastructure.
About

Built by a practitioner.

Sparrow is built and maintained by Andrej Baranovskij, founder of Katana ML and Red Samurai Consulting. It powers document intelligence workflows for enterprise clients in production.

If you need help deploying Sparrow or building document processing pipelines on your own infrastructure — that's exactly what we do.