PII never reaches
your AI.

Veil detects names, amounts, case numbers, and identifiers in any text and replaces them with tokens before anything touches a model. 22 million parameters. Runs on your device. Zero data leaves.

Download Veil
$ pip install veil-phantom
macOS 14+ · Windows 10+ · Free during beta · SDK: Apache 2.0
Shade Live Demo
Scanning...
100%
PII catch rate across 399 entities
22M
Parameters (vs 300M competitors)
<50ms
Inference latency
97.12%
F1 detection accuracy

Every AI tool you use is a data pipeline
you don't control.

  • 1 Client names, deal terms, and patient records processed on servers in jurisdictions you never agreed to.
  • 2 No visibility into what leaves your network. No audit trail. No way to pull it back.
  • 3 Regex and keyword filters miss what matters. Names get mangled by speech-to-text. Context gets ignored entirely.
Third-party cloud Foreign jurisdiction Thabo Molefe R18.5M Absa Ltd Sandton 9402115**** 15 March 2026 Naledi Dlamini +27 82 456 **** Sensitive data from your machine leaving your network Your device Third-party cloud Foreign jurisdiction Thabo Molefe R18.5M Absa Ltd Sandton 9402115**** 15 March 2026 Naledi Dlamini +27 82 456 **** Your device

Three layers between your data and any AI.

Step oneDetect

Shade scans any text: transcripts, documents, chat logs, support tickets. It identifies every name, amount, date, identifier, and entity. Phonetic embeddings catch names that speech-to-text mangles. Regex never could.

Step twoTokenise

Each sensitive value is replaced with a structural token: [PERSON_1], [ORG_2], [AMOUNT_1]. The meaning stays intact. The real data stays on your machine. The AI sees structure, not secrets.

During the meeting, Thabo Molefe presented the R18.5 million offer from Sasol to acquire the Sandton office. Naledi Dlamini from Absa raised concerns about the March 2026 deadline.

Hover highlighted text to see tokens

Step threeRehydrate

After the AI responds, Veil maps tokens back to real values on your device. Summaries, action items, and insights read naturally. The AI never knew who was in the room.

Meeting Summary

Discussed Sasol acquisition and due diligence timeline.

ACTION ITEMS
Trained on 72 million words. 862,000 examples. On a GTX 1050 Ti.

Every conversation. Searchable forever.

Veil builds a private knowledge graph from your meetings. People, topics, projects, and connections surface automatically. Context compounds over time. Nothing forgotten. Nothing exposed.

Thabo M. Q3 Revenue Sasol Deal Naledi D. Thabo M. Q3 Revenue Sasol Deal Naledi D.

The app for your team. The SDK for your stack.

Desktop App

Veil App

Record meetings, get AI summaries and action items, search every conversation with Memory Garden. Shade runs automatically. Every name, figure, and identifier is tokenised before anything reaches the cloud.

  • On-device transcription (Parakeet TDT v3)
  • AI summaries via Claude Haiku
  • Memory Garden knowledge graph
  • Memory Chat: query meeting history
macOS · Windows · Free during beta
Download Veil
Open Source

VeilPhantom SDK

Drop three lines of Python into any AI pipeline. Redact PII, call your model with safe tokens, rehydrate the response. Works with OpenAI, LangChain, and anything that takes a string.

  • 7 detection layers, 19 entity types
  • Token-direct with full rehydration
  • 6ms average overhead per call
  • 93.3% tool accuracy in agentic benchmarks
Python · Apache 2.0 · pip install veil-phantom
View on GitHub

Where data exposure ends careers.

Legal
Privilege preserved by architecture, not policy. Client names and case strategy never leave your firm.
Medical
Patient names and health data stripped before any AI processing. HIPAA-ready by design.
Financial
Deal terms, portfolio data, and client identities stay on-premise. Built for GDPR, POPIA, and SEC compliance.
Government
Classified stays classified. On-premise deployment available. Zero external data transmission.

Shade V7. 22M parameters that outperform models 10x their size.

PhoneticDeBERTa: DeBERTa-v2 fine-tuned with Double Metaphone phonetic embeddings. It learns name patterns, not just spellings. When speech-to-text turns "Nkosinathi" into "Ink Casino Thea," Shade still catches it.

97.12%
F1 detection accuracy
22M
Parameters (vs 300M Presidio, 209M GLiNER)
7 layers
NER, gazetteers, NLP, regex, contextual
12 types
Names, orgs, money, dates, IDs, cards, addresses

The privacy layer that should have existed from day one.

Open source. On-device. Zero leakage.

Free during beta. No account required. Apache 2.0.