Structured Data Extraction
& Agentic AIStructured Data
& Agentic AI
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.
Upload a document. Define a schema. Get structured JSON.
Try it live →More than extraction. A document-AI stack.
Three composable pipelines — from structured extraction to instruction calling and autonomous, multi-step agent workflows.
Sparrow Parse
Schema-validated JSON from invoices, statements, forms and tables — Vision LLMs backed by custom table parsing and validation logic.
Sparrow Instructor
Instruction calling beyond extraction — text processing, analysis, field validation and decision-making.
Sparrow Agents
Chain classification, extraction and validation into monitored, multi-step agentic workflows.
Built for control, not convenience.
Local Inference Only
Documents never leave your infrastructure. No cloud API calls, no third-party data routing. Full control.
Pluggable Architecture
Mix Vision LLM, OCR, and agent pipelines. Runs on Apple Silicon, NVIDIA GPU, or CPU via vLLM or MLX.
Enterprise Ready
In production with enterprise clients. Rate limiting, usage analytics, API-first design, commercial licensing available.
Three steps to structured data.
Drop a document
Images (PNG, JPG) or multi-page PDFs. The file is removed after inference completes.
Describe the fields
Provide a JSON schema listing the fields you want to extract and their expected types.
Validated output
Sparrow returns structured, schema-matched JSON. Ready to drop straight into your pipeline.
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.
Agents that reason over documents.
Sparrow chains steps into autonomous workflows — classifying, extracting, validating and deciding, with every run monitored visually.
Identify
Detect the document type and route it to the right pipeline.
Parse
Vision LLM reads the fields; custom logic reconstructs tables and line items.
Check
Validate values against the schema; flag and retry on mismatch.
Act
Apply business logic — like tax code prediction — and hand off clean results.
Works across document types
From simple receipts to complex multi-section financial documents.
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.
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.