"I got a text saying my bank account is blocked. Is it real?" "मुझे अगले हफ्ते की प्रेजेंटेशन के लिए अंग्रेज़ी की प्रैक्टिस करनी है।" "Monday 10 maths lecture, 10:30 dentist. Can you help me plan?"
Auriv understands all three, in the user's own language, and helps them act on each one safely.
Inspiration
Most AI assistants assume that people type fluently, read English comfortably and already know what to ask. Many people do not: family members, the students I teach, and anyone who has received a "verify your account now" message and wondered whether to trust it.
Many people also find today's AI apps tricky to use: long answers, technical language and interfaces designed for confident typists. For a large part of the general public, that is enough to give up.
The aim was not another chatbot but a guide: an assistant that people can simply speak to in their own language, which keeps them safe, plans real tasks with real dates, and teaches them to do things themselves, with answers that are short, organised and easy to act on.
NVIDIA Nemotron made this possible. Its ability to follow a clear structure reliably and to explain things in plain language meant Auriv could give everyone, from a parent checking a suspicious text to a student planning their week, answers they can understand and use straight away.
What Auriv does
Auriv is a voice-first personal AI available in 20 languages, organised around four everyday needs:
- Organise my schedule: converts spoken meetings and deadlines into a dated table, identifies clashes and suggests preparation time.
- Work toward a goal: recognises goals expressed naturally, then drafts and saves them automatically as a checklist spread across the available time, with a reminder only when a deadline is close.
- Learn something: explains concepts (from photosynthesis to MHD helicity), runs language lessons in which a microphone button listens to the learner and checks each word, and supports hands-on skills with links to reliable video guides.
- Stay safe: every message check opens with SCAM, SUSPICIOUS or LOOKS OK and closes with the user's official national reporting route (UK 7726, India 1930).
Lessons, schedules and goals are saved and can be resumed at any time, with the full conversation restored.
Languages and countries
20 languages: English (UK, US and India), Hindi, Bengali, Tamil, French, German, Spanish, Italian, Dutch, Polish, Portuguese (Brazil), Japanese, Mandarin Chinese, Korean, Indonesian, Vietnamese, Turkish and Swahili. Auriv listens, replies and reads answers aloud in the chosen language, and recognises requests in any of them.
10 countries with verified local support: United Kingdom, United States, India, France, Germany, Spain, Italy, Brazil, Japan and Indonesia. For each, Auriv uses checked emergency numbers, official government sources for questions on documents, tax, benefits and health, and, where available, the official scam-reporting route (UK 7726, India 1930 and cybercrime.gov.in, US reportfraud.ftc.gov). Every number is recorded with its official source and the date it was checked.
How it was built
Three NVIDIA Nemotron models on Nebius Token Factory, each assigned to the task it suits best:
- Nemotron-3.5-Lightning: everyday replies, schedules, goals, scam checks and goal drafting
- Nemotron-3-Ultra-550B: complex concepts, comparisons and non-English lessons
- Nemotron-3-Nano-30B: identifying the type of request in any language
With reasoning mode disabled, most replies arrive within 1 to 7 seconds.
Technology: Python and Flask; an orchestrator that routes each message to one of 13 specialised skills; Tavily web search with social media and video sites excluded; SQLite for goals; and a single accessible web interface using the Web Speech API for voice input and output.
Safety enforced in code, not only in prompts:
- Emergencies receive verified local emergency numbers before any AI model is called.
- Official questions are answered from government sources for the user's country.
- Phone numbers that do not appear in the sources are removed, preventing invented helplines.
- Pasted scam messages are never sent to web search.
- Verified recent legal changes take precedence over the model's training data.
Deterministic dates: expressions such as "next Monday", "five days after that", "in three months" and "13 नवंबर" are converted to exact dates by the application, as language models frequently miscalculate weekdays.
Challenges
Calibration at scale. Auriv was evaluated with 99 routing prompts and 65 complete conversations across more than eight languages, and every reply was reviewed individually. The first full evaluation flagged 39 replies; the final evaluation flagged 5, none of which were genuine errors, and every reply was returned in under 7 seconds.
A persistent model behaviour. When asked "can you teach me…" in Hindi, the Lightning model occasionally responded "No, I am a language model made by NVIDIA researchers." Prompting alone could not resolve this, so Auriv now detects the behaviour, switches to the Ultra model, and answers questions about its identity with fixed text in every language.
Natural voice interaction. This required correcting misheard words without drawing attention to them, ensuring the assistant never transcribes its own voice, prioritising the user's typed corrections over the original transcription, and reducing pauses between spoken sentences.
Clear boundaries. Auriv never makes payments, bookings or sends messages on the user's behalf; it explains how to complete each step.
What I learned
- Reliable AI combines models with deterministic code. Dates, emergency numbers, source filtering and quality checks are best handled in code.
- Multilingual support goes beyond translation. It includes grammar consistent with the assistant's voice, familiar English terms in brackets, and understanding the same request expressed in many different ways.
- Systematic review of real outputs is essential. Nearly every improvement came from evaluating actual replies one by one.
Future vision
- More countries and languages: extending verified emergency numbers, official sources and scam-reporting routes beyond the current 10 countries and 20 languages.
- Guided "do it with me" mode: Auriv waits while the user completes each step of a real task, such as an online form or a bank app, and helps when they are stuck.
- Sharing goals with a family member or helper: with the user's permission, a trusted person can see and support the goals they are working on.
- Progress over time: lesson and goal history that shows what the user has learned and achieved, so Auriv can suggest the right next step.
Developed by Rashmi Sarwal for the Nebius x NVIDIA Global AI Hackathon.
Log in or sign up for Devpost to join the conversation.