Developer Builds Bilingual AI Agent with RAG, Voice, Chat and Full Admin Dashboard

·1 views

A developer has built a production-ready AI assistant that handles phone calls, web chat, WhatsApp, and Instagram through a single shared knowledge base using Retrieval-Augmented Generation (RAG). The system supports both Hindi and English, automatically detecting and responding in whichever language the user writes or speaks. A diagnostic admin panel exposes retrieval scores, KB hit rates, latency, and error metrics without requiring a separate analytics pipeline. Lead capture works conversationally, parsing names, emails, and phone numbers from free-text or voice input and assigning them to sales reps via round-robin rotation. Multi-tenant branding is handled through runtime variable interpolation rather than separate deployments, and the data layer can be switched from SQLite to Postgres via the admin UI without redeploying the backend.

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)

Why AI Agents Keep Repeating the Same Errors — and How to Fix It

AI agents commonly fall into retry loops when tool calls fail, often because error messages like '422 validation_error' provide no actionable context to determine whether a retry is worthwhile. The core problem is not the model's reasoning but the absence of structured failure memory that agents can consult before attempting the same action again. A useful failure trace requires three components: a normalized error identity, a recorded outcome rather than just an intention, and a success denominator to make failure rates meaningful. Normalizing raw error strings by stripping timestamps, UUIDs, and request IDs allows separate incidents to be recognized as the same underlying bug. The open question remains whether failure patterns across different users' agents actually overlap enough to make shared error knowledge viable — a hypothesis the author says is testable but currently unresolved.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

VibeAudio Plays Music During AI Agent Waits, Chimes When Task Is Done

VibeAudio is a new open-source Node.js tool designed to eliminate the frustrating idle period users experience while waiting for AI coding agents to finish tasks. The tool plays procedurally generated background music during agent processing and plays distinct audio cues to signal success, failure, or a required user prompt. Audio tracks are synthesized in real time without any audio files, and the style of music adapts based on how deep or complex the agent's current task is. Each working directory generates a consistent musical fingerprint, so the same project always produces the same sound pattern. VibeAudio supports popular AI coding tools including Claude Code, Cursor, Codex, and Gemini CLI, and can be installed globally via npm from its GitHub repository.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Safely Roll Back an AI-Generated Pull Request You Cannot Explain

Developers working with AI coding agents are advised to follow a structured rollback workflow when a pull request becomes difficult to understand or justify. The process begins with pausing CI runs, marking the PR as a draft, and notifying reviewers before taking any further action. Engineers are encouraged to generate a local file inventory of all changed paths using git diff commands, saving the output for methodical review rather than reading patches line by line. A decision table guides the choice between reverting and fixing forward, recommending revert when changes fall outside the ticket scope, tests are failing, or a developer cannot explain a specific code hunk. The workflow is intended as a personal onboarding aid rather than team policy, and is specifically designed for developers who still own the branch and are not operating under a required merge queue.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Flawed Netgear Routers Overwhelmed UW-Madison Time Server in 2003

In 2003, the University of Wisconsin-Madison's internet time server was flooded with excessive traffic caused by buggy Netgear routers. The routers contained a firmware flaw that caused them to repeatedly query UW-Madison's NTP time server at abnormally high rates. This design defect meant that as more affected routers were deployed, the volume of requests to the server grew unmanageable. The incident highlighted how a widespread consumer hardware bug could inadvertently overwhelm a public internet infrastructure resource. Details of the problem were documented by UW-Madison researcher David Plonka on the university's computer sciences website.

0 comments Read more at Hacker News