AI Enterprise SolutionsFinalistGPT & Claude

Pərdə — Local Privacy Layer for Azerbaijani AI Workflows

Pərdə is for customer-support operators at Azerbaijani banks, telecom companies and government service centres. Their customer messages contain names, FIN codes, identity documents, card numbers, phone numbers and addresses. AI could help summarize complaints, classify tickets and draft replies, but pasting the original messages into an external chatbot exposes that personal data. Operators must therefore redact messages manually or avoid AI assistance. Pərdə detects and masks personal data locally before sending a request to an AI model, then restores the original values in the answer locally. It supports Azerbaijani-specific identifiers, name morphology and mixed Latin/Cyrillic text. On 1,000 held-out Azerbaijani texts, it protected 94.9% of labelled personal values, compared with 23.9% for our generic regex baseline. Detection is not perfect; operators can inspect the masked request before sending it.

Open the demo Video GitHub

Preliminary score

84.0/ 100Strongaverage of 2 judges

Value for the user22.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI23.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility13.0 / 15
Data requirements, running costs and a clear next step.
Originality7.5 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 62.0 / 100 on stage.

What the judges wrote

Claude Opus 5.5Preliminary83/100
Strengths
Your user and pain are specific: support operators in Azerbaijani banks, telecoms and service centres who cannot safely paste tickets with FIN codes, cards and names into an LLM. Your evaluation is unusually rigorous for a hackathon. It uses a held-out split of a public Azerbaijani PII dataset, a regex baseline, an ablation showing the name detector's contribution (41.2% → 94.9%), per-entity results, CI reproduction and an honest list of failures and fixed bugs. You built Azerbaijani-specific detection covering case endings, patronymics, mixed scripts and checksummed IDs. The end-to-end mask → local Gemma → restore loop is demonstrated, along with a hash-only audit log.
To improve
Most of your own engineering is deterministic rules, and the LLM is a generic downstream step. A small local Azerbaijani NER model as a second detector would strengthen the AI contribution and lift address (82.6%) and date (80.3%) recall. Your real-model evidence is a single integration sample. Run the full task flow (summaries and draft replies) across dozens of tickets and report tag-preservation and restoration failure rates. Validate on even a small set of real, anonymised support tickets, since the benchmark is machine-translated. Measure operator time per ticket so the business metric goes beyond leak rate.
GPT-6 AstraPreliminary85/100
Strengths
Your local masking-and-restoration workflow addresses a concrete barrier to AI adoption for Azerbaijani support teams. Your held-out evaluation, name-detector ablation, entity-level results and explicit failures provide strong evidence for the language-specific detection work. The reported Gemma integration establishes a real-model path beyond the clearly labelled mock demo, while your deployment instructions make the prototype practical to reproduce.
To improve
Your evaluation should report strict full-value masking alongside the current 90%-character threshold, since partial masking can still disclose information. Expand testing to securely handled real support tickets and more real-model interactions, measuring answer usefulness, omitted or invented tags, and performance against an established multilingual PII tool. Your pilot should compare operator time against manual redaction, quantify inference cost per ticket, and establish when residual privacy risk requires blocking external transmission.

The team

Jeyhun NaghiyevLead
data
Ruslan Qasimli
developer
Mubariz Aslanov
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: any OpenAI-compatible
Data: Public - LocalDoc/pii_ner_azerbaijani (CC BY 4.0)
Components: streamlit, openai SDK, pytest
Built during the event: Confirmed

Quality testing

Today’s practice: staff paste customer messages into a chatbot as they are, so 100% of the personal data reaches the AI provider. Main test: 1,000 held-out Azerbaijani texts from the public LocalDoc PII dataset (CC BY 4.0). They contain 2,065 labelled personal values. We never used this split for tuning; we tuned on 2,000 separate dev texts. Results (share of personal values protected): Pərdə: 94.9% (leak rate 5.1%), precision 92.8% Generic regex filter: 23.9% (leak rate 76.1%) Pərdə without its Azerbaijani name detector: 41.2%, which shows the name handling makes the difference By type: phones, e-mails, ID numbers (FIN, ID card, passport, VÖEN) and cards 100%; names 97.4%; addresses 82.6%; dates 80.3% Other tests: 20 hand-written messy tickets mixing Azerbaijani, Russian and English: 98.7% protected. We wrote these ourselves, so we treat them as a smoke test only. End to end with a real model (Gemma 4 E4B on a local llama.cpp server): all tags were kept and the real values were restored correctly. 20 unit tests pass. CI builds and runs the Docker image, and reproduces the held-out numbers on every push. What broke: Rare names written in Cyrillic with no surname next to them, such as "Ульвия". Patronymics joined to a suffix, such as “İsmayılqızıdır”. The date format 1987/01/01, which we left unfixed so the test score stays honest. Some ordinary words that are also names get masked (“Ümid”, “Ulduz”). That’s the safe direction, but it’s why precision is 92.8%. A code review also found real bugs, which we fixed: personal data typed into the instruction field was sent unmasked; the leak check raised false alarms; codes like ABC1234 were masked as FIN numbers. Limitations: The dataset is machine-translated, so real tickets may score lower. The leak blocker cannot catch values the detector missed.

Worth 20 of the 100 points.

Running it

Browser demo: Open the demo link. No installation or credentials are required for offline mock mode. Select a sector to load a sample, inspect the detected values and masked text, then click “Protect and send.” Mock mode demonstrates masking and restoration without calling a real AI model. Local setup on Linux/macOS: git clone https://github.com/Wkrotik/gamesummitai2026.git cd gamesummitai2026 python3 -m venv .venv source .venv/bin/activate pip install -r requirements-dev.txt streamlit run app.py Open http://localhost:8501. On Windows, with Python 3.10 or newer installed, double-click run.bat instead. To use a real local model, start an OpenAI-compatible model server and select “llamacpp” in the app. Set the model name to one served by your server. The default endpoint is http://127.0.0.1:8080/v1; override it using PERDE_LOCAL_BASE_URL if needed. We tested gemma-4-E4B on llama.cpp without an API key. External providers require their respective API keys.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.