II before text reaches any LLM.

The problem: AI agents processing meetings, emails, support tickets are handling raw sensitive data. Names, salaries, medical details — all flowing through cloud APIs.

The solution: Detect PII on-device, replace with tokens ([PERSON_1], [AMOUNT_1]), send safe tokens to LLM, rehydrate response locally.

Interesting finding: In benchmarks (98 scenarios, 8 verticals, Claude Haiku), accuracy went UP with PII redaction — 91.5% → 93.3%. Token-structured input seems to help models parse arguments more reliably.

Technical details: - Shade V7: 22M param PhoneticDeBERTa (DeBERTa-v3-xsmall + Double Metaphone embeddings) - Trained on 72M words of meeting/business data - 7 detection layers (NER, gazetteers, regex, NLP, contextual) - 19 PII token types - 6ms average overhead - 97.1% F1 on meeting transcripts

The phonetic embeddings help catch ASR-mangled names across cultures — "Nkosinathi" transcribed as "Ink Casino Thea" still gets detected.

pip install veil-phantom

Docs + benchmarks: https://helloveil.com/sdk GitHub: https://github.com/helloveil/veil-phantom

Apache 2.0. Happy to answer questions about the architecture or training approach.19:T605,With vibe coding tools, adding multilingual support has become surprisingly easy. If you ask a model to add languages like Spanish, Portuguese, German, or French, it can usually set up the i18n structure pretty quickly, even for projects with a lot of text.

One thing I’ve noticed is that once a site is indexed by Google, traffic doesn’t always come mainly from the US. Sometimes a significant portion comes from other countries through search. But most launch platforms or directories are English-only, so they don’t really help with discoverability in other languages.

Because of that, I tried something d

FFmpeg-over-IP – Connect to remote FFmpeg servers

An Alternative to Play Integrity

Revise age verification terms for MidnightBSD

Claude Code Skills for Startup Founders – 12 Commands for Strategy, Not Code

Scott Aaronson's View of My View About Quantum Computing

More tokens, less cost: why optimizing for token count is wrong

Meta acquires AI agent social network Moltbook

Time Travel Has Been Done

Qt SQL Under LGPL Despite MariaDB Under GPL

Thinking Deeply About Theming and Color Naming

Codex Monitor

Show HN: VeilPhantom – Open-source on-device PII detection for AI pipelines

10 years of AlphaGo: The turning point for AI [video]

NASA and SpaceX disagree about manual controls for lunar lander

Ask HN: Does automatic multilingual support make sense for a launch platform?

Planning a multi-city trip without the spreadsheet nightmare

Designing Interfaces for AI Agents Instead of Humans

The Clarity Act, Stablecoin Yield and the Chinese Crypto Bogeyman

Show HN: OpenMolt – A programmatic AI agent framework for Node.js

AI, Human Cognition and Knowledge Collapse (NBER Working Paper)

Agent API Spec Design: When API Callers Change from Application to AI Agent

Prioritizing energy intelligence for sustainable growth

US added more solar than any other technology in 2025, but is down 14% from 2024

Cardiac side effects of RNA-based SARS-CoV-2 vaccines

Epstein Fallout Tracker

Why AI is both a curse and a blessing to open-source developers

Astro 6.0

Fixing request smuggling vulnerabilities in Pingora OSS deployments

Using cookies to hack into a tech college's admission system

StackOverflow Site Redesign Thread (2026)

Ask HN: How will AI-generated code be vetted?

Comments