Developer Builds Multilingual Voice AI Finance Assistant for Rural India in 10 Days

·1 views

A developer built RupeeGPT, a voice-first AI financial assistant, in 10 days to help Indian users navigate banking, UPI, government welfare schemes, and loans through natural conversation. The tool supports English, Hindi, and Hinglish, targeting first-generation bank account holders, rural farmers, and street vendors in tier-2, tier-3, and rural areas. The system uses Deepgram for speech recognition, Google's Gemini for language processing, and Murf Falcon's Indian English TTS voice to deliver a culturally familiar experience. A custom pronunciation layer converts Roman-script Indian scheme names into Devanagari before text-to-speech output, ensuring authentic pronunciation of terms like PM Kisan Samman Nidhi. The project was built using Python, LiveKit, Next.js, and MongoDB Atlas, and was developed as part of a 10-day voice agent challenge.

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 Shares Playwright Solution for Visual Testing with API Mocking

A developer has published their solution to Challenge 3, focused on visual testing combined with API mocking using Playwright. The approach addresses real-world problems such as missing chart renders and overlapping UI buttons caused by frequent API changes in dashboard applications. The solution involves mocking API endpoints to return fixed data, then capturing and comparing screenshot snapshots to detect unintended UI changes. A custom VisualHelper class was built with methods to check full-page or element-level snapshots, with configurable timeout and pixel-difference thresholds. Playwright's built-in toHaveScreenshot() function serves as the free alternative to paid visual testing tools like Percy, Applitools Eyes, and Sauce Visual.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer recreates kulhad chai in pure CSS for frontend art challenge

A developer built a detailed CSS illustration of a kulhad of masala chai for DEV Community's Frontend Challenge: Comfort Food Edition. The artwork depicts a hand-thrown terracotta cup surrounded by whole spices and a half-eaten biscuit, rendered entirely using divs, gradients, clip-paths, and shadows — no SVG, images, or canvas. The kulhad's irregular silhouette was achieved with an 18-point polygon clip-path, while five stacked gradient layers simulate the texture of matte clay. Steam wisps required three iterations to look realistic, ultimately using individual horizontal origins and mix-blend-mode: screen to mimic rising vapour. A CSS mask-image technique was used to punch a clean bite-shaped hole through the biscuit, preserving its crumb texture and shadow details.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DEV Community Post Shares CSS Code for Arabic-Themed Restaurant Landing Page

A developer published a front-end project on DEV Community showcasing a landing page concept called 'Comfort Table,' styled around an Arabic home-cooking theme. The post consists almost entirely of raw CSS and HTML code rather than written editorial content. The design uses a warm, earthy color palette with serif typography to evoke a cozy, home-style dining atmosphere. The project appears to be a personal or portfolio exercise demonstrating responsive web design techniques. No additional context, author background, or publication date was provided in the source material.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Compares Rule-Based and ML Anomaly Detection in Custom Log Analysis Tool

A software developer added an Isolation Forest machine learning model to Log Sentinel, a self-built Apache log analysis dashboard originally designed to detect brute force attacks and directory scans using rule-based logic. The experiment used a labelled dataset of 230 IP addresses — 200 normal and 30 simulated attackers — to fairly evaluate both detection approaches. Rule-based detectors achieved perfect precision and recall because the evaluation data was deliberately structured around known attack patterns the rules were written to catch. The ML model matched on recall by catching all 30 attackers but produced five false positives, flagging normal IPs with statistically unusual behaviour. The developer concluded both methods are complementary: rules reliably catch known threats with no false alarms, while ML can surface unexpected anomalies that no predefined rule would cover.

0 comments Read more at DEV Community