AI Document Pipeline

What's in the public demo

The public demo is a browser-based version of the AI Document Pipeline — launching Summer 2026. You open a PDF on your own computer and the demo runs locally, with nothing leaving your device. Free to use, no account required.

• Runs in your browser — The demo runs entirely on your computer using modern browser features (with automatic fallback on older browsers). The AI model downloads once on first use; after that, the demo works offline.
• Nothing leaves your device — Your PDFs and your analysis stay on your computer. After the AI model downloads, you can disconnect from the internet and the demo keeps working. The privacy promise is something you can verify yourself.
• Mark documents as private — Tag a document as confidential and its source text is hidden from the answer-generation step — answers reference your knowledge graph instead. You can share insights without revealing the underlying confidential text.
• EU AI Act transparency disclosures — Every AI-generated output is clearly marked, in line with EU AI Act Article 50. You can see exactly what was AI-generated versus what's from your source material.

What's in the engagement deliverable

When you engage Synavistra, your team takes home a complete, working AI pipeline tailored to your documents. Apache 2.0 licensed — you own everything and can keep using it after the engagement ends, with no ongoing fees or vendor lock-in.

• Take-home Open Text Bundle — A complete bundle from each engagement: a knowledge graph of the entities and relationships in your documents, structured answers with source citations, and a full audit trail — all in open, plain-text formats your team can inspect, share, or modify.
• Custom-trained AI model (inference-offline variant) — For the inference-offline engagement variant, you also receive a Gemma 4 model fine-tuned on your domain documents, plus the training code and infrastructure recipe. Run the AI on your own hardware indefinitely — air-gap capable for sensitive material.
• Reproducible by design — Every export is content-addressed: the same source documents through the same recipe produce byte-identical results. That's what makes the audit trail genuinely verifiable, not just claimed.
• EU AI Act compliance package — Every bundle ships with a machine-readable compliance file and a human-readable compliance summary, plus a full article-by-article assessment of the EU AI Act. Hand it to your auditors as-is.

See what the product looks like before you engage — every showcase is a finished example built on a public document.

Two products, two engagement paths

Which product you receive depends on the engagement variant. Both deliver the full methodology applied to your domain documents.

• Inference-online
  Anthropic Claude, Google Gemini, or Mistral APIs process your documents during the engagement on Synavistra accounts. You take home the recipe (Apache 2.0), a complete log of every interaction with the AI, a signed provenance record, and instructions to set up your own API account if you want to keep running the analysis afterwards.
• Inference-offline
  Everything in the inference-online product, plus a customer-fine-tuned Gemma 4 model, the training-stack source, and the recipe to reproduce the full training infrastructure on your own cloud or hardware. Runs offline indefinitely on your own hardware, air-gap-capable for sensitive data.