# Redpill Linpro CX

- **Event:** [RAISE Summit Hackathon](https://cerebralvalley.ai/e/raise-summit-hackathon)
- **When:** Jul 4 at 9:00 AM – Jul 5 at 7:00 PM (GMT+2)
- **Where:** Paris, France
- **Team:** [Selim Sevim](https://cerebralvalley.ai/u/selimsevim)
- **GitHub:** https://github.com/selimsevim/clipboard-firewall
- **Demo video:** https://www.youtube.com/watch?v=4t3IY7WRFbU
- **Gallery:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/79

Marketers and CRM consultants paste customer data into AI chats every day. Emails, personnummer, customer IDs, names inside SQL literals. Each paste is a governance incident waiting to happen, and manual redaction is slow and error-prone.
Clipboard Firewall is a macOS menu bar app that sanitizes your clipboard on-device, before the data can leave. Copy a customer export, a support ticket, or an SFMC query; by the time you paste, PII is replaced with consistent pseudonyms ({{PERSON_1}}, {{CUSTOMER_ID_1}}), while code, SQL schema, and public information stay untouched. Clean text passes through silently. One click restores the original.
Under the hood: a hybrid detector. A deterministic regex layer with checksum validation handles fixed formats (email, phone, IBAN, personnummer, credit card). A fine-tuned Gemma 4 E4B, running fully locally via Ollama, handles what patterns cannot: names in mixed Swedish/English text, coreference ("Anna Lindqvist", "Anna", "she" map to one pseudonym), quasi-identifiers ("our only customer in Kiruna over 90"), and customer IDs that depend on context. The model only detects; verified code performs every replacement, so it can never rewrite your text.
Built during the hackathon: the full app, a 190-example synthetic dataset, and a LoRA fine-tune (trained on a cloud GPU, merged to GGUF, shipped on-device). On the frozen held-out set, customer ID recall went from 69% to 100% (13 of 13) with zero false positives on clean text and 100% valid JSON. The failed training attempts and fixes are documented in the repo.
No cloud calls, no telemetry, one endpoint: localhost. Turn off Wi-Fi and it keeps working, because nothing it does ever needed the internet.

## More from RAISE Summit Hackathon

- [Face to face](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/76)
- [BlackoutGuard](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/77)
- [Xeam | Medyt](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/78)
- [CoVibers](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/80)
- [SHTC](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/81)
- [DriftGuard](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/82)

---

Markdown version of https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/79. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt