SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Building Guardrails So an LLM Can't Corrupt a Production ERP Database

0
·5 views

The author's team integrated a low-cost LLM into a production ERPNext system through an MCP gateway. In testing, the model fabricated a successful database query even though its tool calls had never executed, showing it can invent results when tools are unavailable. Rather than relying on smarter models or better prompts, they built a design in which the assistant is read-only and writes require a separate executor account and explicit user confirmation. Each action is recorded in an audit ledger, and the system is verified through census and scenario tests. The guarantees cover the main MCP path and enumerated entities, while side channels and session-level risks are excluded.

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 ·

Build a Typed AI Support Triage Feature in .NET 10 Using IChatClient

The article shows how to add a basic AI feature to a .NET 10 app without a vector database or agent framework, relying on the IChatClient interface. It walks through creating a minimal API project and installing the Microsoft.Extensions.AI packages along with the OpenAI adapter. The author recommends storing API keys in user secrets and registering the chat client once through dependency injection. A support ticket triage endpoint is built that returns a typed C# result containing a category, priority enum, and summary. The piece also covers streaming reply drafts and swapping models with a single change.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How to Fix Core Web Vitals on a Next.js and Sanity Site, Starting With Mobile

A Next.js and Sanity site showing Poor or Needs Improvement URLs in Search Console can be triaged by focusing on one failing metric at a time. The guidance stresses that Search Console reflects real-user field data over a rolling 28-day window, so a strong Lighthouse score does not offset it. Mobile is usually the first place failures appear and should be addressed first. Pages are grouped by template, so fixing one shared template can clear many URLs at once. The article outlines the 75th-percentile thresholds for LCP, INP and CLS, and covers common causes such as lazy-loaded hero images, missing reserved image dimensions, font swaps and heavy client-side JavaScript.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Bot-Generated Articles Skew Wikipedia Word Frequencies in Low-Resource Languages

An analysis of word frequency lists from several Wikipedia editions suggests that some minority-language versions are unreliable as linguistic datasets. In balanced human-written text, function words like prepositions and conjunctions dominate, but Tatar and Bashkir Wikipedias are topped by content words such as "river" and "water" because many articles were auto-generated from templates using state water registry data. Erzya and Moksha Wikipedias show a similar distortion, with frequent terms tied to bot-created orchid stubs, a pattern the author calls the Orchid Phenomenon. Average article lengths in these editions are also much shorter than in Russian Wikipedia, raising the risk of repetitive template-driven content. The author advises developers to profile corpora with simple frequency counts before training language models on Wikipedia dumps.

0 comments Read more at DEV Community