A challenge to document 30 entries of building, detecting, and defending with AI in trust & safety. Paced as I have time. Living index — updated as projects ship.

I’m running a challenge: 30 entries of leveraging AI in Trust & Safety, in public. Every project I ship, every experiment I run, every wiki entry I write gets a page — and they all link from here.

This isn’t a daily streak. Entries land as and when I have time — some weeks two or three, some weeks none. The “30” is the count of substantive things I want to ship and document, not consecutive days. The point is to anchor focused work on one question: how do we leverage AI in T&S when fraudsters are already using it really well? Some entries are build logs, some are deep dives on detection solutions, some are syntheses of papers I’ve been reading.

This page is the living index. It grows as entries land.

The shape of it

• Project 1 — ScamDecoder: AI-powered scam detection for consumers, live at scamdecoder.com.
• Project 2 — FraudDecoder Brain: an LLM-maintained wiki of fraud modus operandi and prevention controls.
• Project 3 — LLM Email Fraud Detector: an LLM-driven detector for suspicious emails.
• Project 4 — Synthetic Fraud Data for Testing: generating synthetic fraud data to stress-test detection systems.

Index

Each project has its own page; build logs, experiments, and lessons live under it.

Project 1 — ScamDecoder

AI-powered scam detection for consumers, live at scamdecoder.com.

• ScamDecoder — project page
• What v2 looks like — coming soon (revamp with Claude design)

Project 2 — FraudDecoder Brain

An LLM-maintained wiki of fraud modus operandi and prevention controls.

• I built an LLM-maintained wiki for fraud — kickoff
• The schema doc as the project — coming soon
• The Telegram approval loop — coming soon
• Mechanism over metrics: the quality bar — coming soon

Project 3 — LLM Email Fraud Detector

An LLM-driven detector for suspicious emails.

• Building an LLM email fraud detector with LangChain — coming soon (next article)

Project 4 — Synthetic Fraud Data for Testing

Generating synthetic fraud data to stress-test detection systems.

• Generating synthetic fraud data for testing T&S systems — coming soon

Why this challenge

Three reasons:

1. The field is moving too fast for passive reading. Every week brings new MOs, new defenses, new regulator action. Building forces me to actually understand, not just skim.
2. AI is a forcing function in T&S specifically. It’s simultaneously the new attacker (deepfakes, autonomous fraud agents), the new defender (LLM detection, agentic controls), and the new substrate (wikis, eval harnesses, copilots). Most other domains only get one of those at a time.
3. Public commitment beats private intent. Writing this index out loud is the cheap version of accountability.

How to follow along

• This page updates as projects and posts land — bookmark it.
• Daily-sized posts go on LinkedIn.
• Longer writeups land here on the blog under the Trust & Safety thread.

— Sofia