A new voice-first assistant, Arthashathi, evolving from a hackathon challenge, seeks to make financial and public-service information more accessible to Indian users through real-time, multilingual voice interaction and specialised routing.
Arthashathi began as a simple challenge project and quickly grew into something more ambitious: a voice-first financial guide for Indian users that tries to make complex information easier to access through conversation. In a detailed write-up on dev.to, the builder said the aim was to move beyond a basic talking chatbot and create an assistant that could help people ask about government schemes, banking safety, fraud concerns and general financial questions without having to sift through websites or dense terminology.
The project reflects a practical problem in India, where useful financial and public-service information is often available but not always easy to navigate. Arthashathi was designed with that in mind, especially for users who are more comfortable speaking than typing. It supports real-time voice interaction, memory for returning users, code-mixed and Indian-language communication, and a set of safety rules intended to keep the assistant away from sensitive details such as one-time passwords, PINs, CVVs and bank credentials.
Under the hood, the system uses LiveKit for real-time audio transport and Murf Falcon for text-to-speech. Murf’s own integration guide and LiveKit’s documentation both describe the same core promise: low-latency voice synthesis and agent orchestration for conversational applications. The Arthashathi build combines those tools with speech recognition, language-model reasoning, memory, escalation logic and analytics, showing how a production-ready voice agent depends on more than just an LLM answering questions.
One of the most interesting parts of the project is its focus on specialist routing. Rather than forcing one general-purpose agent to handle every query, Arthashathi can hand off scheme-related conversations to a dedicated government-scheme specialist. That design choice matters because it reduces the burden on the main agent and keeps each component narrowly focused. The builder said the lesson was simple: a good multi-agent system is not about quantity, but clear responsibility.
The project also explored outbound calls, human escalation and call-level analytics. Those additions point to a broader ambition than a traditional assistant: not just reacting to users, but proactively reaching them where appropriate, while still leaving room for a person to step in when the AI reaches its limits. The author also described several technical setbacks, including a dispatch problem caused by an agent-name mismatch and latency issues that made the assistant interrupt too quickly when silence detection was too aggressive.
The final result is less a finished product than a working blueprint for how voice AI can be built carefully. The Arthashathi journey suggests that the hardest part of voice systems is not any single model or tool, but the orchestration between them: speech recognition, routing, state, safety, voice output and recovery when something breaks. That is also why the project’s larger takeaway is so practical. Start small, the builder advises, then add complexity only after the basic speech-to-speech loop is reliable.
For now, Arthashathi remains a work in progress, with plans for better multilingual support, smarter routing, stronger failure recovery and broader real-world testing. But even in its current form, it shows how a focused build can turn a hackathon-style challenge into a serious prototype for financial assistance that is more conversational, more local and more useful than a standard text box.
Disclaimer: This article is intended to inform and educate, not to recommend or endorse any financial product, investment or strategy. Please consider your own financial circumstances and seek professional advice where appropriate before making financial decisions.
