AI Models Card

This single card covers the three Dedomena.AI EDGE Engine packages that distribute pretrained AI models: dedomena-edge-text, dedomena-edge-vision and dedomena-edge-audio. It describes what those models do, how they are constrained in production, and where they fail.

1. Scope

Dedomena.AI EDGE ships two very different kinds of AI, and only one of them is a distributed model.

2. Model details

3. Execution profile

All models execute inside the customer’s own Google Cloud project, on the customer’s GPU host. Content being analysed never leaves that project.

Inference is deterministic by configuration: sampling temperature is set to zero for all detection and verification steps, so the same input yields the same findings.

Model weights are delivered with the Engine package. There is no runtime download of weights from a third party and no call-out to a model API.

The speech-recognition task runs quantised on CPU; the vision and language stages require the GPU.

No public ingress is opened in the customer project. The only outbound traffic is licence validation and usage counters.

4. dedomena-edge-text

4.1 What it does

Locates personal data and named entities in documents and free text so they can be redacted, masked or substituted with synthetic data. It is a detection and rewriting tool, not a general-purpose assistant. Allows to generate natural language insights.

4.2 Models used

4.3 Limitations

• Detection is probabilistic. False negatives (personal data left in place) and false positives (over-redaction) both may occur, human revision of the risk analysis is always recommended.

• Performance varies by language, domain and document layout. Scanned, handwritten, or poorly structured documents are materially harder than clean digital text.

• Subtle nuance, sarcasm and figurative language are unreliably interpreted, which matters where identifying information is implied rather than stated.

• Biases and gaps in the underlying training data propagate into detection coverage — entity types and name forms that are under-represented upstream are detected less reliably.

• The underlying model can generate text that is incorrect. Substituted values are plausible, not verified; they must never be read back as facts.

5. dedomena-edge-vision

5.1 What it does

Locates identifying elements in images and video — faces, signatures, logos, fingerprints and text-bearing regions such as identity-document fields — and removes them by regenerating the affected region so the result stays visually coherent rather than being covered by a black box. Face and fingerprint detection is performed solely to locate potentially identifying elements and apply a privacy transformation to them, and the Engine does not perform biometric identification, biometric verification, emotion recognition or biometric categorization of people.

5.2 Models used

5.3 Safety mechanisms specific to this Engine

The pipeline is built on a separation of authority that is worth stating explicitly, because it is the property that makes an editing model acceptable in a privacy tool:

> The detector never decides. It only proposes. Every proposed region is confirmed or rejected by a separate verification step before anything is modified.

• Regional editing only. The editor is constrained to the confirmed region; the rest of the image is untouched, so the document's structure, background and unrelated content cannot drift.

• Multi-pass additive detection. Images are analysed more than once — direct, contrast-equalised for faint elements, filtered for microtext, and tiled for small regions — and results are fused by overlap. Passes only ever add findings, never remove them.

• Machine-readable zones are erased and redrawn, never rewritten in place, so no residue of the original encoded values survives.

• Objective acceptance thresholds. A redaction is accepted only if it measurably worked: text fidelity is scored against the requested replacement, and residual detail and residual contrast in an obscured region must fall below fixed limits. Failing output is rejected, not shipped.

• Deterministic fallback. When no generated candidate meets the thresholds after all seeds are exhausted, the pipeline falls back to a deterministic obscuring operation. The failure mode is a visibly cruder image, never an image that looks anonymized but is not.

• Coordinates confirmed once are reused downstream rather than re-detected, because two detectors never agree pixel-for-pixel and a second detection pass would leave uncovered slivers.

5.4 Limitations

• An unanalysed image cannot be assumed clean. Absence of findings is not evidence of absence of personal data.

• Performance varies with resolution, occlusion, motion blur, lighting and compression, and may vary across demographic groups.

• Microtext, low-contrast overprints and heavily stylised or rotated text are the hardest cases and the most likely source of a miss.

• Regenerated regions are synthetic content. A processed identity document is a fabricated image of a document and has no evidentiary value.

• This is the most computationally demanding Engine; large images and video are slow and GPU-memory bound.

6. dedomena-edge-audio

6.1 What it does

Transcribes speech, locates personal data in the transcript, and produces anonymized audio and transcript output.

6.2 Models used

6.3 Safety mechanisms specific to this Engine

The governing risk here is upstream of the anonymization step:

> A transcription error propagates. What does not appear in the text is not detected, and therefore is not anonymized.

• Language is detected before transcription rather than assumed, because a wrong language assumption degrades the transcript and silently degrades detection with it.

• Decoding is primed with the identifier formats of the detected language, which materially improves the capture of exactly the strings that need removing.

• Values that cannot be anchored to a timestamp in the audio are recorded explicitly rather than dropped, so the operator knows a span was found but could not be located for removal.

6.4 Limitations

• Accuracy varies with recording quality, background noise, overlapping speakers, accent and dialect.

• Control characters and check letters in spoken identifiers are frequently lost, and word boundaries are approximate, so timestamp-aligned removal can clip or overrun by a fraction of a second.

• Identifier priming is currently configured for Spanish, French and English. Other languages transcribe but without format priors, and detection of identifiers in them is correspondingly weaker.

• Running the recognition stage quantised on CPU bounds throughput. Long recordings are slow.

• Speaker identity may remain inferable from voice characteristics even when every spoken identifier has been removed.

7. Intended use

7.1 Intended uses

• Removing personal and sensitive data from unstructured content so it can be analysed, shared or retained under a lawful basis that the original content did not support.

• Preparing training, test and demonstration material without exposing production personal data.

• Reducing the manual effort of redaction under human supervision.

7.2 Intended users

Data, privacy, risk and compliance teams and data engineers, operating inside an organisation that holds the source content lawfully, and competent to review the output.

7.3 Out-of-scope uses

• Re-identification of individuals, or any attempt to reverse anonymization.

• Presenting a processed document, image or recording as a genuine original. Output is synthetic and must never be used as evidence, as proof of identity, or to obtain a service.

• Use of the Engines as general-purpose generative models. They are constrained to the detection and anonymization pipeline described here.

• Sole or automated basis for decisions with legal or similarly significant effects on individuals.

8. Data handling

9. Customer-trained synthesizers (context, not a distributed model)

The core dedomena-edge package provides proprietary generative algorithms for tabular, transactional, time-series and relational data. No trained model is distributed: the customer trains a synthesizer on their own data, in their own project, and owns the result. It is included here only because its limitations are commonly asked about alongside the Engines.

• Fidelity depends on the source data. Small or sparse datasets may produce weaker synthesizers.

• Undeclared cross-field business rules may not be preserved; declared relational integrity is.

• Synthesization reduces disclosure risk but does not mathematically eliminate it. Membership- and attribute-inference remain possible in principle, particularly for synthesizers trained on small and simple datasets.

• Built-in evaluation (synthesizer_evaluations and synthesizer_evaluations_multitable) measures utility and privacy before output is used, and the risk-detection function scores re-identification risk before and after protection.

10. Fairness and bias

• Detection quality is not uniform across groups. Face detection can vary across skin tones and image conditions; speech recognition varies across accents, dialects and languages; entity detection varies across name forms and scripts. Uneven detection means uneven privacy protection — the people least well served are those whose data is most likely to be missed.

• The Engines inherit socio-cultural bias present in the pretraining data of the underlying open-weight models.

• Identifier priming in the audio Engine currently favours Spanish, French and English, which is a known and explicit fairness gap rather than an incidental one.

• Customers with fairness-sensitive use cases should measure subgroup performance on their own content rather than relying on aggregate figures.

Dedomena.AI has not published quantitative subgroup evaluation for the Engines: there are no measured detection rates broken down by skin tone, accent, dialect, language, script or document type. Two things follow. First, the pipeline is built so that uneven detection is contained rather than hidden: it over-detects and then verifies, every redaction is scored against objective acceptance thresholds, output that fails them is rejected and reprocessed by a deterministic fallback, and nothing is released without human review — so a subgroup weakness tends to surface as a visible failure an operator can catch rather than as content that looks protected and is not. Second, customers whose use case is fairness-sensitive should measure performance on their own content and their own population, since a vendor-published aggregate would not answer that question for their data in any case. Per-modality and per-language subgroup evaluation is a committed roadmap item, and results will be published in a future revision of this card.

11. Evaluation

The Engines are evaluated against objective acceptance criteria applied per item at run time, not only against a benchmark at release. Redactions are scored for text fidelity, residual detail and residual contrast, and output that does not meet the thresholds is rejected and reprocessed by the deterministic fallback path. Synthesizers are evaluated on the customer's own data, which is the only distribution that matters for their use case.

12. Ethical considerations

Three principles govern how these models are operated, and they are stated here because they explain design decisions elsewhere in this card.

> A false negative is a leak of personal data. The pipeline is therefore tuned to over-detect and then verify, rather than to detect conservatively and under-cover.

> Generative models are used to remove or replace information. Processed output is synthetic by construction and is marked and intended as such.

> Automated anonymization supports human judgement and does not replace it. The product produces material for review; it publishes nothing and takes no action on the customer's behalf.

13. Legal status of output

Anonymization by these Engines is not a legal determination that the output is anonymous data under the GDPR or any other regime. Anonymity is context-dependent and assessed against the data, the recipient and the means reasonably likely to be used for re-identification. Dedomena.AI provides detection, removal, replacement and measurement; it does not make that determination. Customers remain the data controller and should obtain their own legal advice. It is up to the client to determine their role, legal basis and other regulatory obligations applicable to the treatment.

14. Recommendations for safe use

• Sample-check output manually in every new document type, image source or recording condition before trusting the pipeline on that class of content.

• Treat any unprocessed item as unprotected. Do not infer coverage from the absence of findings.

• Review risk against the intended recipient rather than in the abstract — the same content carries different risks shared internally and published.

• Keep a human reviewer on any release into a regulated or high-risk context.

• Stay on a supported version. Corrections to model behaviour and detection coverage are delivered through package updates.

15. Version and change history

Model Card Version: 1.0.2

Last Update: Aug 19th, 2026