EdgeHome Harness Treats 1B AI Model as Untrusted Input to Control Smart Devices

·2 views

A community developer built EdgeHome Harness, a 25 MB Rust framework pairing with OpenBMB's MiniCPM5-1B model to create a reliable smart-home controller on devices with just 2 GB of RAM. The design deliberately limits the AI model's role to generating a candidate JSON command, while deterministic Rust code handles all validation, device resolution, capability checks, and policy enforcement. The harness explicitly labels the model's output as 'untrusted', treating it the same way a web application treats raw user input. Under low-memory conditions, the system progressively shrinks the model's context and output budget, and can drop the model entirely in favor of rule-based fallbacks. The project argues that agent reliability is a function of harness design rather than model capability, with every execution plan running as a dry run before deployment.

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)

Developer Builds AI-Powered Legal Management System for Zimbabwe Law Firm

A software developer in Zimbabwe built MutemoOS, a custom AI-assisted legal operating system, to help his wife's Harare law firm manage over 118 active cases previously tracked in a diary and notepad. The system integrates Anthropic's Claude, ChromaDB, PostgreSQL, and live Zimbabwe legislation APIs to handle case research, document drafting, and deadline tracking. The developer found that mainstream legal AI tools are designed for large firms in the US or UK and lack knowledge of Zimbabwean-specific legal concepts, courts, and statutes. To address AI hallucination risks, he built a two-stage pipeline that first checks whether retrieved sources are sufficient before synthesising answers, flagging unsupported claims for lawyers to verify. A companion service scrapes Zimbabwean legal databases daily to keep the system's knowledge current across multiple law firm clients.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Codebase Structure, Not Prompts, Determines AI Coding Agent Performance

AI coding agents like Cursor and Claude Code perform significantly better when the repositories they work on follow clear structural conventions, according to a developer analysis. Key factors include organizing code into folder-per-concern layouts with one component per file, rather than large monolithic files that force agents to reason about entire codebases at once. Explicitly typed and exported component props act as a readable API, allowing agents to use components correctly without guessing at implementation details. Configuration files such as .cursorrules and CLAUDE.md provide persistent system-level context to agents on every run, reducing the need to re-explain project conventions in each prompt. These structural choices are described as low-cost to implement but capable of substantially improving the reliability and output quality of AI-assisted development.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Overusing useState Is Slowing Your React Apps, Developers Warned

A developer guide published on DEV Community highlights a common React anti-pattern where too many component values are stored in useState, causing unnecessary re-renders. The article identifies three main culprits: derived values that should be computed with useMemo, values only used in event handlers that belong in a useRef, and form fields that rarely need to be controlled inputs. The author argues that useState should only hold values React genuinely needs to track for rendering, not every variable in a component. Alternatives like useRef, closure variables, and reading form data on submit are recommended for cases where re-renders add no benefit. The core takeaway is that avoiding unnecessary state is a more effective performance strategy than memoizing expensive renders after the fact.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Agents Are Replacing Workers, Not Just Assisting Them, Tech Leaders Warn

Microsoft AI CEO Mustafa Suleyman predicted in February 2026 that artificial intelligence would reach human-level performance on most professional tasks within 12 to 18 months. Salesforce CEO Marc Benioff revealed in September 2025 that AI agents already handle roughly half of all customer service interactions at his company, with its support workforce cut from 9,000 to around 5,000 employees. The shift marks a fundamental change from AI as a productivity tool to AI as an autonomous agent capable of completing entire workflows without human involvement. A KPMG survey from Q2 2025 found that 33 percent of organisations had already deployed AI agents, up threefold from the previous survey period. Analysts and commentators warn that the resulting labour disruption will be faster and broader than the gradual automation scenarios policymakers spent years preparing for.

0 comments Read more at DEV Community