Radar / Ideas / Airgap Extract: audit-ready document…

Airgap Extract: audit-ready document intelligence that never leaves the building

daily ideaambitiousJEV confidence 0.572026-10-02
Outcomeevery loan file, KYC packet, and contract converted to cited, confidence-scored structured data with zero bytes leaving the customer's infrastructure

The problem

Banks, insurers, and hospitals want agentic document AI, but the best extraction products are cloud-only APIs - uploading customer records to a vendor's servers is a compliance violation under RBI data-localization rules, GDPR, and HIPAA before the first token is generated. Self-hosted alternatives exist but lack citations and per-field confidence, so compliance teams cannot sign off on their output.

The idea

A single deployable pipeline that renders Office documents offline with a self-contained LibreOffice engine, runs an agentic extraction harness (structural reasoning, verification passes, bounding-box citations - the pattern LlamaIndex Extract v2.5 proved) using open models inside the customer's VPC, and gates every extracted field through a small typed decision model returning calibrated confidence. Fields below threshold route to a human review queue with the source highlighted. Output is validated JSON with an immutable audit log recording what was extracted, by which model, at what confidence.

Why now

Three missing pieces landed at once: DeepSeek's dsh-libreoffice-kit gives agents an offline Office rendering engine with bundlable binaries; LlamaIndex Extract v2.5 published the harness recipe (structural reasoning plus verification plus citations, value F1 up from 89.8 to 95.8) that can be replicated with open models; and open decision models (kev family, Strands Decider 2B) provide the cheap confidence gate. India's DPDP Act enforcement and the EU AI Act's August 2026 high-risk obligations make 'the cloud API saw our documents' an increasingly expensive answer.

What it combines

Offline Office rendering (deepseek-dsh-libreoffice-kit) + agentic extraction harness techniques (llamaindex-extract-v2-5) + typed decision gates (kev-open-trainable-jev-like-family-of-small-decision-models-on-qwen3-5-3-8). The kit renders but does not understand; the extraction harness understands but is cloud-only; decision models judge but need a pipeline to judge inside. Combined they produce something none offers: extraction quality with citations and confidence, running entirely behind the customer's firewall.

MVP

Weekend scope: Docker container that takes DOCX/XLSX/PDF, renders via the offline Office kit, extracts a loan-application schema with an open vision-language model, attaches bounding-box citations, and runs per-field confidence with a review queue for anything under threshold. Deliberately skip: handwriting, 130+ format coverage, multi-page table stitching, and custom model training.

Distribution

B2B2C: embed as the extraction engine inside loan-origination and core-banking platforms, including Indian banking SaaS vendors serving co-operative banks that regulators note are largely shut out of AI adoption. The platform vendor sells AI document processing to its bank customers; the startup licenses per deployment or per document. Who pays: banks, NBFCs, and insurers processing KYC, loan, and claims paperwork under data-residency mandates.

Why it wins

LlamaIndex Extract, Azure Document Intelligence, AWS Textract, and Rossum are all cloud APIs - documents must be uploaded to their servers at $0.005-$0.075 per page. Open self-hosted tools like Struktur keep data local but ship no citations and no confidence scores. Airgap Extract is the first to combine self-hosting with cited, confidence-gated output and a human review queue - the exact properties a bank's risk team needs to approve it.

Risks

Biggest risk is an extraction-quality gap versus hosted agentic tiers on messy layouts. The MVP de-risks by benchmarking against a labeled set of real loan documents and leading with the confidence gate: the product promise is that humans review only uncertain fields, which holds even if raw accuracy trails the cloud state of the art.

Build it with

Repo to start from

airgap-extract - offline document-intelligence pipeline: Office rendering plus agentic extraction with citations plus calibrated confidence gates, zero network egress.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.