Ozulon Privacy Shield

Company: Ozulon Privacy Shield

Self-reported basis: Author-supplied project description from a Devpost submission for the OpenAI 2026 hackathon

Commercial due-diligence read: This is a self-reported, pre-launch AI-native desktop application designed to locally detect and redact sensitive data in screenshots and video screen captures. The product is described as a proof-of-concept or early-stage prototype with no evidence of revenue, customers, or market traction. It uses local AI inference with a fine-tuned DeBERTa model and claims performance metrics for recall, precision, and latency. The author states that the system is not yet production-ready and is iterating on real-world scenarios.

Key open question: Is there any indication that this product will become viable as a commercial offering, or whether it has moved beyond a hackathon prototype?

The description states that Ozulon Privacy Shield is a desktop application that detects sensitive data in screenshots and video screen captures. It operates locally on the user’s device and proposes redaction tracks for video.

• Screenshot functionality: Detects sensitive information, allows users to review findings, and apply protection (Mosaic, Blur, Solid).
• Video functionality: Analyzes source video using adaptive sampling, proposes time-bounded redaction tracks, and allows user review before export.
• Feedback loop: Users can provide local feedback that improves future detection without uploading data; feedback is protected with DPAPI.
• Model & tech stack: Uses a fine-tuned DeBERTa-v3-xsmall v6 INT8 model (~87 MB), ONNX Runtime, OCR, and Microsoft DeBERTa.
• Runtime environment: Runs locally, not in the cloud. GPT-5.6 was used for research but is not part of the shipped runtime.

Inference: The product appears to be a desktop tool built for developers or teams that share screen captures, with an emphasis on privacy and local processing.

The author positions Ozulon Privacy Shield as a way to make “local AI redaction on screen captures the default,” aiming to improve safety in screen sharing workflows.

• Original inspiration: Screen captures expose sensitive data (names, credentials, addresses) and are used for teaching, explaining, and shipping work.
• Evolution of claim: The product evolved from an earlier deep-scan approach that was too resource-intensive into a faster, local solution suitable for everyday use.
• Key claim: It makes privacy protection “a natural part of creating and sharing screen-based work,” without requiring users to send sensitive content to the cloud.

Inference: The positioning is focused on developer or team workflows where screen sharing is common, and privacy is a concern. The product is framed as a tool that improves default behavior rather than a standalone service.

The description does not clearly identify a specific customer segment or ideal customer profile (ICP).

• Implicit use case: Teams or individuals who frequently share screen captures in work environments.
• Possible users:
  • Developers or engineers sharing code or UI designs
  • Teams documenting processes or training materials
  • Anyone using screen capture tools for internal or external communication

Inference: The target is likely a niche segment of professionals or teams who are sensitive to privacy and use screen sharing as part of their workflow.

There is no evidence in the description of a business model, pricing strategy, monetization plan, or customer acquisition approach.

• Self-reported: The product is described as pre-launch.
• No revenue data, pricing tiers, or customer contracts are mentioned.
• No indication of whether it will be sold as a SaaS, freemium, or one-time purchase.

Inference: No business model is evident. It may be a prototype or early-stage product with no commercialization strategy yet defined.

The author provides technical details about how the product was built and what it does:

• Model: Fine-tuned DeBERTa-v3-xsmall v6, INT8, ~87 MB
• Performance metrics:
  • 95.87% end-to-end recall
  • 94.73% strict precision
  • ~100 ms mean detector latency after OCR
  • ~300 MB peak memory use
• Video scan time reduction: 85.2% reduction from 178 to 26.33 seconds for a two-minute video
• Delivery stack:
  • Built with .NET 8, C#, FFmpeg, GitHub Actions, Inno Setup
  • Uses ONNX Runtime, Microsoft DeBERTa, OCR, OpenAI Codex (for research only)
  • Installer includes verified local model, tokenizer, configuration, FFmpeg, FFprobe, and desktop runtime

Inference: The technical stack is mature for a desktop application. The product shows engineering sophistication in model selection, performance optimization, and local processing.

There is no evidence of traction or maturity beyond the hackathon prototype:

• Pre-launch status: Described as “pre-launch” and “just beginning.”
• No customers, revenue, or adoption data.
• Bugs acknowledged: The author states that there are bugs and that evaluations do not yet represent real-world scenarios.
• Iterating on feedback: The team plans to expand coverage, improve video performance, and refine controls.

Inference: The product is in an early stage of development and has not yet reached a market-ready or customer-facing state.

The description does not mention competitors or the broader competitive landscape.

• No direct competitor names are listed.
• No positioning relative to existing tools for screen redaction or privacy protection.
• No mention of how this compares to other local or cloud-based solutions.

Inference: No evidence of competitive analysis or awareness of the market. The product is described in isolation, without reference to similar offerings.

Several risks and red flags are evident from the self-reported description:

• Pre-launch status: Product is not yet commercialized or tested with real users.
• No revenue or customer data: No evidence of traction or monetization strategy.
• Unverified performance claims: Metrics like recall, precision, and latency are self-reported without independent validation.
• Limited team size: Only one member (the founder).
• GPT-5.6 used for research only: No indication that AI is being used in production or at scale.
• No clear path to market: No mention of go-to-market strategy, distribution, or user acquisition.

Inference: The product is a prototype with no commercial viability demonstrated. It lacks the evidence needed to assess whether it will become a viable business.

1. What are the actual use cases you’ve tested this with?
2. How do you plan to monetize or scale this product beyond the hackathon prototype?
3. Have you validated the performance claims (recall, precision, latency) in real-world usage?
4. What is your go-to-market strategy for reaching users who share screen captures regularly?
5. Are there any legal or compliance considerations around local data handling and feedback loops?
6. How do you plan to handle internationalization and cultural variations in sensitive data?
7. What are the key technical challenges that remain unresolved before shipping to production?