SShortSingh.
Back to feed

Travel Site Uses LLM for Moderation but Lets Humans Have the Final Say on Negative Reviews

0
·45 views

A developer building 'Back From My Trip', a travel review site, has designed an AI moderation pipeline that explicitly protects negative reviews from being automatically rejected. The system assigns every piece of content a moderation status — approved, needs_review, or rejected — with row-level database security ensuring only approved content is publicly visible. While the AI can approve content autonomously, it cannot issue a final rejection for text-based opinions; any AI-flagged rejection is routed to a human admin queue for confirmation or override. When the moderation system fails or exceeds its budget, content defaults to needs_review rather than auto-approval, ensuring failures never allow spam through or silence a legitimate opinion. The developer also patched an early security flaw where client-side text could be swapped before moderation, replacing it with a server-side trigger system that processes only stored row IDs.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Microsoft Ignite 2026: Copilot Repositioned as Work Platform, With Agent and Model Updates

At Microsoft Ignite 2026 in San Francisco, CEO Satya Nadella opened by addressing last week's global Microsoft 365 outage that disrupted Outlook and Teams, then described Copilot as the new operating system for work. The rebuilt Copilot brings Chat, Cowork, Code, and Autopilot into one experience, with Home and Office generally available now. Several headline features, including shared Copilot Workspaces, the long-running Autopilot agent, and Copilot Code, are rolling out through a Frontier preview program or remain not yet generally available. Microsoft also made its Decision 1.1 model generally available, positioned for classification and routing between agents, and released Code 2 models on GitHub Copilot and Azure AI Foundry, with performance and cost claims that Microsoft has not independently verified.

0
ProgrammingDEV Community ·

Keeping related-link blocks stable across publishing platforms

A DEV Community post discusses how to keep related-link blocks consistent when content is published on multiple platforms. The author recommends generating the link block once and appending it after the article body. The reason is that platforms vary in how they support markdown and HTML. The post also references a source repository, a Mbin instance, and the author's DEV profile.

0
ProgrammingDEV Community ·

Detecting Privilege Escalation Through Enumeration Bursts in Linux Logs

A DEV Community article describes a method for spotting attackers during the reconnaissance stage, before they escalate privileges. It notes that a service account running several enumeration commands within a few minutes is unusual in normal operations. The author proposes counting distinct enumeration command types per user ID per minute and alerting when three or more appear, excluding root. In testing on a synthetic log, the detector flagged a burst from one account using four command types while ordinary single- or double-command admin activity stayed silent. The author recommends allow-listing known inventory jobs by user ID and source to reduce false positives, and says a Windows version is planned.

0
ProgrammingDEV Community ·

Fake bank apps installed via callers or texts pose growing fraud risk, warn researchers

Scammers may phone or text people and urge them to install a counterfeit bank app to resolve a supposed account problem. Visa's Spring 2025 threats report cautioned users against downloading apps from potentially malicious sources. On March 30, 2026, Poland's CERT described a tactic in which a bank-branded app prompted users to install an extra component and enable accessibility settings. Experts advise installing banking apps only from official app stores or from a bank's website typed in manually, and ending any call that guides you through an installation.