GPT-4's Hindi output is barely functional for nuanced medical queries. Ask any native speaker to describe a "heavy stomach" or a "dull ache" in their mother tongue, then ask an English-trained LLM to translate or reason with it. The gap is not just wide; it's a chasm. This isn't a dig at the incredible progress in large language models; it's an observation about the fundamental bias in AI training data - heavily skewed towards English and a few other major global languages. This specific, often-ignored problem is precisely what we're solving at GoDavaii, India's Advanced Health AI.
My name is Pururva Agarwal, and as a 27-year-old founder from India, I've spent the past few months (and many years observing the problem) understanding why generic AI falls short for families here. The biggest engineering and product challenge isn't just the sheer volume of medical knowledge, but making that knowledge accessible and contextually relevant in every language an Indian family might speak.
Beyond Simple Translation - Why Context is King
The immediate assumption for multilingual AI is often: translate English content into other languages. But health isn't that simple, especially not in a culturally rich and linguistically diverse country like India. The nuances of describing a "khansi" (cough in Hindi) - dry, wet, chesty - are often conveyed through regional idioms and common parlance, not clinical terms.
Our AI Health Chat in 22+ Indian languages goes far beyond dictionary lookups. We're training models to understand these implicit meanings, cultural references, and regional variations in symptom description. This means building incredibly precise linguistic models for each language, often starting with limited data, and then cross-referencing that understanding against a vast medical knowledge graph. It's about recognizing that health advice given to an aunty in Indore needs to sound different, feel different, and address different concerns than to a professional in Bengaluru, even if the underlying medical condition is similar.
Architecting for a Multilingual Health Graph
To power features like our Drug Interaction Checker or the Lab Report AI explanation, we've had to rethink how medical information is structured. Instead of an English-centric knowledge base with translation layers, we're building a multi-modal, multi-lingual health graph from the ground up. This involves ingesting medical information, drug data, and health guidelines directly in various Indian languages whenever possible, and then building intelligent bridges between these knowledge silos. For instance, understanding how common terms like "पित्त" (pitta), which in common parlance often refers to acidity or indigestion, relates to clinical gastrointestinal conditions requires deep contextual understanding, not just a lexical match.
This architecture is vital for accuracy. When our system identifies a potential drug interaction, it's not just checking an English database and translating the warning. It's reasoning through the context of the medicines, their active ingredients, and common usage patterns as understood in the local language. This is a technically demanding path, far from the easier route of simply localizing an existing English product. It's a testament to why we placed Top 14 Global Finalist at Startup Flight Vietnam 2025 - this underlying language stack is genuinely unique.
Bridging Traditional and Modern - The 'Desi Ilaaj' Challenge
Another core moat for GoDavaii is our AI-verified Desi Ilaaj. How do you use AI to verify traditional home remedies? For generations, remedies like "haldi doodh" (turmeric milk) for a cough or specific herbal concoctions have been part of Indian households. Our goal isn't to replace modern medicine, but to provide families with AI-verified context on these practices: what's the scientific basis (if any), what are potential interactions with allopathic medicines, and when should a doctor be consulted instead? This requires a delicate balance of cultural respect, scientific rigor, and linguistic precision.
It's a huge challenge, ethically and technically, to build AI that understands cultural practices without dismissing them outright, or conversely, without endorsing unverified claims. We're developing models that can cross-reference Ayurvedic principles with modern pharmacological data, all while conversing naturally in multiple Indian languages. It helps families raise the questions that matter of their doctors, bridging the information gap with cultural relevance.
We're building for families who think and speak in languages AI has largely ignored, ensuring that health information is not a privilege for English speakers. GoDavaii is designed to be a preparation tool for families, helping them surface questions to ask their doctor, in their own language, ensuring they get the answers that matter. What linguistic challenges are you tackling in your own projects? Drop your thoughts in the comments below.
Top comments (0)