G-Sharp

G-SHARP™

GDPR-compliant AI with zero context leakage

We pseudonymize before any model sees input—so people stay private and data stays useful.

G-SHARP™ – Protect the person. Preserve the data.
European patent application filed (EPO 25206504.0)

How it works (high level)

G-SHARP™ stepwise transforms raw text into placeholders, then reassembles a fully pseudonymized version. Context is stripped before any AI step, so models only ever receive de-contextualized word lists—not the original text. :contentReference[oaicite:1]{index=1}

Known inputs (e.g., names, IDs) are replaced first. Pattern-matched entities (addresses, dates, IDs) follow. Typos are handled by fuzzy matching. :contentReference[oaicite:2]{index=2}

We then split text into tokens, remove frequent words to remove context, and shuffle what remains. AI detects only GDPR-relevant tokens from that randomized list (first names, last names, etc.). :contentReference[oaicite:3]{index=3}

Finally, an iterative check step validates and fixes edge cases, and the text is reconstructed with placeholders. :contentReference[oaicite:4]{index=4}

Step {{StepNumber}} — Detect Language

The system quickly checks likely languages to configure subsequent steps.

English
Spanish
German

Die schulische Leistungsentwicklung der Schülerin Suzanne Fischer (SchulID 5W1773, Schwester von Jonas Fisher aus der 10B), geboren am 3. Juli 2007 und wohnhaft in der Göthestraße 12, 79100 Freiburg …

Step {{StepNumber}} — Replace Known Inputs

What happens (1–2 lines):

All provided known inputs are replaced with placeholders (e.g., “Suzanne” → [Vorname:1]). :contentReference[oaicite:5]{index=5}

Die schulische Leistungsentwicklung der Schülerin Suzanne[Vorname:1] Fischer[Nachname:1] (SchulID 5W1773, Schwester von Jonas Fisher aus de 10B), geboren am 3. Juli 2007, wohnhaft in der Göthestraße 12, 79100 Freiburg …

Step {{StepNumber}} — Pattern-matched Entities

Addresses, dates, IDs matched by regex/patterns become placeholders (e.g., “Göthestraße 12” → [Straße:0], “3. Juli 2007” → [Datum:0]). :contentReference[oaicite:6]{index=6}

… geboren am 3. Juli 2007 [Datum:0] , wohnhaft in der Göthestraße 12 [Straße:0] , 79100 [Postleitzahl:0] Freiburg …

Step {{StepNumber}} — Fuzzy Match

Detect likely typos (e.g., “Fisher” ≈ “Fischer”) using Levenshtein and normalize to known inputs. :contentReference[oaicite:7]{index=7}

… Schwester von Jonas Fisher [Nachname:1] aus der …

Step {{StepNumber}} — Tokenization

Break up the text into a list of individual word tokens.

Die schulische Leistungsentwicklung der Schülerin [Vorname:3] [Nachname:1] ( SchulID 5W1773, Schwester von Jonas [Nachname:1] aus der 10B), geboren am [Datum:0] und wohnhaft [Straße:0], [Postleitzahl:0] in Freiburg …

Step {{StepNumber}} — Remove Frequent Words

Common function words and punctuation are removed since they are not PII but do provide context.

Die schulische Leistungsentwicklung der Schülerin [Vorname:3] [Nachname:1] ( SchulID 5W1773 , Schwester von Jonas [Nachname:1] aus der 10B ) , geboren am [Datum:0] und wohnhaft in der [Straße:0] , [Postleitzahl:0] Freiburg

Step {{StepNumber}} — Insert Decoy PII Tokens

Add realistic decoys to the list as an active protection layer before AI labeling.

schulische Leistungsentwicklung Schülerin SchulID 5W1773 Schwester Jonas 10B geboren wohnhaft Freiburg

Step {{StepNumber}} — Shuffle Remaining and Decoy Tokens

Shuffle the remaining and decoy tokens to remove residual context before AI processing.

schulische Leistungsentwicklung Schülerin SchulID 5W1773 Schwester Jonas 10B geboren wohnhaft Freiburg Michael Rutgers 030-3643981

Step {{StepNumber}} — Iterative AI on De-contextualized Tokens

AI receives only the shuffled list and converts all GDPR-information to placeholders, e.g., “Jonas” → [Vorname:4], “10D” → [Identifikation:2])

Schwester wohnhaft Rutgers 5W1773 schulische 030-3463981 Michael SchulID 10D geboren Schülerin Jonas Freiburg Leistungsentwicklung

A.I.

Step {{StepNumber}} — Iterative Checks

From minimal context, AI corrects placeholders (e.g. 10B is not an identifier but a class) and fixes misses (e.g., catch Zürcher as a last name). :contentReference[oaicite:11]{index=11}

… aus der [Identifikation:2] [Klasse:1] ). …

Step {{StepNumber}} — Remove Decoy PII Word Tokens

We’re starting to piece things back together. The decoy PII word tokens that were added earlier have served their purpose of hiding context from AI. They are removed.

Schwester wohnhaft [Nachname:5] [Identifikation:1] schulische [Telefonnr:2] [Vorname:3] SchulID [Klasse:1] geboren Schülerin [Vorname:4] Freiburg Leistungsentwicklung

Step {{StepNumber}} — Restore Original Order

Restore the word tokens to their original order.

Schwester wohnhaft [Identifikation:1] schulische SchulID [Klasse:1] geboren Schülerin [Vorname:4] Freiburg Leistungsentwicklung

Step {{StepNumber}} — Re-insert Previously Removed Word Tokens

All words, placeholders and punctuation that were previously removed are re-insterted to get back to the original list of word tokens, but now fully pseudonymized.

Die schulische Leistungsentwicklung der Schülerin [Vorname:3] [Nachname:1] ( Schwester von [Vorname:4] [Nachname:1] aus der [Klasse:1] ) , geboren am [Datum:0] und wohnhaft in der [Straße:0] , [Postleitzahl:0] Freiburg

Step {{StepNumber}} — Reassemble Pseudonymized Text

Concatenate tokens back to text with placeholders only, ready for compliant AI processing & later re-identification inside the tenant. Presto! Done! :contentReference[oaicite:12]{index=12}

The pseudonymized text can be used as GDPR-compliant AI input.

Die Schülerin [Vorname:3] [Nachname:1] (SchullID [Identifikation:1], Schwester von [Vorname:3] [Nachname:1] aus der [Klasse:1]), geboren am [Datum:0]und wohnhaft in der [Straße:0], [Postleitzahl:0] Freiburg …