Intarsia - local pseudonymization for business documents

What the company appears to be: Intarsia is a self-reported local-first tool for pseudonymizing sensitive business documents. The author states it detects personal names, organizations, IDs, contact details, and more in text-based files (.txt, .md, .docx, .pdf), proposes pseudonymized versions, and allows human review of findings before export. It runs entirely locally on the user’s machine, with encrypted mappings stored only on the user's device.

What changed: The author reports that during a hackathon (Build Week), they used Codex and GPT-5.6 to accelerate product design, UI implementation, QA, and packaging — effectively building a full product solo. This was not an incremental improvement but a new product launch from scratch using AI assistance.

Single most important open question: Is there any evidence of actual usage or adoption beyond the author’s own development work? The description states no revenue, customers, or traction data exist beyond what is self-reported.

Note: All claims in this summary are based on the author's own write-up and are unverified. No third-party corroboration exists for any aspect of the project.

The description states that Intarsia is a local-first workspace for pseudonymizing sensitive professional documents. It accepts text-based formats such as .txt, .md, .docx, or text-based .pdf files. The tool detects people, organizations, contact details, projects, contract and tax IDs, IBANs, credentials, and more.

It proposes a fully pseudonymized version next to the original document. Users can review findings at three levels:

• Accept or reject a whole category (e.g., “Contact, place & time — 42 hits”)
• Review normalized terms (e.g., “Beacon Ridge Migration — 2 hits”)
• Review one exact occurrence

Users may add their own known values and apply changes. Every instance of the same value gets replaced with a stable typed placeholder like <PERSON_0001>. A mandatory re-scan checks results, and critical residuals trigger a visible "review required" state instead of silent completion.

The encrypted mapping stays local so controlled reversal remains possible when explicitly needed.

Claim: Intarsia is a local-first tool for pseudonymizing business documents.

Evidence: Author’s own write-up

The author positions Intarsia as a solution to the problem of AI-assisted analysis of legal and contract documents being blocked by data protection laws. They state that existing tools either over-pseudonymize (destroying usefulness), under-pseudonymize (missing recurring values), or make cleanup so tedious that no one uses them.

Intarsia is described as a tool that offers strong automatic detection followed by a fast, review-first workflow — avoiding the pain of manual pseudonymization while maintaining document integrity. It is optimized for German-language material today but supports English documents through the same workflow.

The name "Intarsia" comes from inlay work: precise pieces fitted into a larger whole so the overall picture stays intact.

Claim: Intarsia solves the problem of AI-assisted analysis being blocked by data protection laws.

Evidence: Author’s own write-up

The author states that consulting, legal, and procurement teams across Europe are the target audience. These teams face a dilemma: contracts would benefit from AI-assisted analysis or drafting, but data-protection obligations demand careful handling before documents with names and identifiers go near cloud AI tools.

In practice, this means teams either pseudonymize manually at high cost, skip AI value, or accept risks they don’t want to own. Intarsia aims to provide a better option: local pseudonymization with automatic first-pass detection, human review of every finding, and stable placeholders that keep documents readable and analyzable.

Claim: Target customers are European consulting, legal, and procurement teams.

Evidence: Author’s own write-up

Not evidenced. The description does not mention any pricing model, monetization strategy, or business model details.

Finding: No evidence of business model or pricing structure.

The tool is built with:

• Local FastAPI backend
• React/TypeScript review workspace
• Optional macOS shell using SwiftUI/WKWebView
• Python sidecar
• Microsoft Presidio Analyzer for deterministic recognizers
• Two pinned GLiNER2 detector snapshots (fastino/gliner2-privacy-filter-PII-multi @ e5b59ed0…, fastino/gliner2-multi-v1 @ cc151f5b…)
• Fernet encryption for mappings in product mode
• Runtime defaults to offline mode after one-time public-model provisioning step that never sends document text

The author reports using Codex and GPT-5.6 to build UI designs, implement workflows, run browser validation, automate packaging, and manage QA.

Claim: Intarsia is built with local-first architecture, deterministic detection, and human review.

Evidence: Author’s own write-up

Not evidenced. The description does not provide any data on usage, adoption, revenue, or customer base. It only mentions that the author is a solo developer who shipped the product during a hackathon.

Finding: No evidence of traction or maturity beyond development phase.

Not evidenced. The description does not name competitors or describe the competitive landscape.

Finding: No evidence of competitive context.

• Solo builder risk: The project is built by a single person (Matthias Olzmann). There is no indication of team structure, ongoing support, or scalability.
• No traction or adoption: No evidence of real-world usage or customer feedback.
• Unverified claims: All descriptions are self-reported and unverified; there is no third-party validation of functionality or performance.
• Limited scope: The tool is optimized for German documents and English documents run through the same workflow, but no indication of broader language support or localization plans.
• AI dependency: Reliance on Codex and GPT-5.6 for development raises questions about whether the product can be maintained without continued access to those tools.

Inference: The lack of any external validation or usage data suggests a high risk of failure in real-world deployment.

1. What is your plan for scaling beyond solo development?
2. Have you tested the tool with actual users from target industries (legal, consulting, procurement)?
3. How do you intend to monetize this product if at all?
4. Are there any known limitations or edge cases in detection accuracy that could affect usability?
5. What are your plans for expanding language support beyond German and English?
6. Can the tool be integrated into existing workflows or systems used by target customers?
7. How do you ensure long-term maintenance of model dependencies like GLiNER2?

Note: These questions are based on the self-reported nature of the description and aim to probe areas where evidence is lacking.