About the client
Our client is a fast-growing quick-service restaurant chain operating across 100+ branches in India and the USA, delivering through various delivery partners such as Zomato and Swiggy in India. The chain is on a 10X growth trajectory heading toward 1,000+ branches in the coming year.
At that scale, the challenge is not just revenue, it's operational knowledge. Recipes, SOPs, and preparation standards need to reach every kitchen, every shift, every new hire, without friction. The challenge was that this knowledge was sometimes needed immediately, mid-task, while actively working and accessing an LMS or a documentation portal in that moment was simply not practical. The solution needed to deliver that information through a voice interface on a mobile device, the moment staff needed it.
Operational impact
Knowledge Under 3 seconds
Instant access, any shift
Knowledge that used to require a search now takes three seconds. Staff ask, the system answers.
Best Experience
Works in your language
Not English. Not standard Hindi. Speak Hinglish or a mix of Hindi or Marathi, no problem.
High Accuracy
Right answer, No hallucinations
Preparation standards remain consistent regardless of who is working. Every cook follows the same source of truth.
Why Google Docs and LMS was not enough for kitchen staff
With growing scale of operations, quality and consistency are the key to business growth. For staff to deliver consistent quality, they need to follow standard operating procedures and to do that, they need the right knowledge quickly and reliably.
The team started with Google Docs. Recipes, preparation guides, and operational procedures were all documented and shared. But it was not the right solution. Staff working in a kitchen during a live service rush are not in a position to open a document and search for an answer.
They moved to a Learning Management System next. The LMS worked well for one-time onboarding and upskilling, structured training that staff could go through before they started a role. But it was not built for frequent, on-the-job access. When a cook mid-shift needs to know something specific, an LMS is not where they turn.
Another challenge is that the frontline staff rotates frequently, which leaves institutional knowledge scattered across documents, WhatsApp messages, and the memory of people who may no longer be there. New hires had no reliable way to get answers during a live kitchen rush. Over time, this leads to inconsistency in how food is prepared across branches.
The team set out to fix this, by making knowledge as accessible as a conversation.
What is a multilingual voice AI assistant?
A multilingual voice AI assistant is a conversational AI system that lets users speak in their natural language including mixed or regional dialects and receive accurate, context-aware responses in real time. Unlike standard voice bots that follow scripted decision trees, a voice AI assistant understands intent, handles follow-up questions, and responds in natural language.
For QSR operations in India, this means kitchen staff can ask questions mid-shift in Hindi, Tamil, Telugu, Kannada, Marathi, or Bengali or a mix of all of them and get answers sourced directly from the chain's own recipes, SOPs, and training guides. No typing, no searching, no waiting for a senior person to be available.
Example: How much paneer do I need for making 300 grams of Butter Paneer Masala? if asked mid-shift in Hindi, answered instantly from the recipe on file.
Figure 1: Sage AI voice assistant answering recipe and SOP queries in Hindi during a live kitchen shift
How it works for QSR kitchen operations
Documents like PDFs, DOCX files, or directly from Google Drive are uploaded once. The system processes, indexes, and understands them automatically. Any updates to existing documents are reflected across the knowledge base without manual intervention.
Why multilingual support is non-negotiable for Indian QSR operations
India is one of the fastest-growing markets for quick-service restaurants, with a diversified culture and dialect landscape. This comes with a challenge, staff speak different languages depending on where they are from, not where they work and in practice, they mix them mid-sentence. That is just how people naturally talk.
Most voice tools are built around a single working language, usually English or standard Hindi. That works for a narrow slice of the workforce. For kitchen staff across multiple cities, a solution that only handles one language is not really a solution.
For this to work, the assistant needed to handle the way people actually speak with regional accents, mixed vocabulary, no expectation that the user adjusts to the tool. That shaped the model choice from the start.
demo: Sage AI multilingual voice agent demo, real-time recipe query and response for QSR kitchen staff
Why we chose Sarvam BulBul v3?
We evaluated multiple voice models and selected Bulbul v3 as the best fit for this use case. Below are the factors that made the difference:
- Latency: A voice assistant that pauses before responding is of no use, staff will just ask a colleague instead. BulBul v3 includes a low-latency streaming mode that generates and plays audio in near real time, which is what makes the interaction feel like a conversation rather than a query tool.
- Built for Indian speech: Most voice models are trained in English and extended to other languages. BulBul v3 is built from the ground up for Indian languages. In an independent third-party blind listening study across 11 languages, it outperformed ElevenLabs and Cartesia on naturalness at telephony grade, the audio quality that actually matters for real-world deployment, not studio conditions.
- Stability: In production, small errors compound. A mispronounced ingredient, a skipped word in a recipe, these are not minor issues, they affect the quality what gets prepared. This model has the lowest word skip rate (7.47%) and mispronunciation rate (7.84%) among the models evaluated, which matters when the system is answering questions across 300+ recipes at every branch, every shift.
- Code-mixing: Kitchen staff across India do not speak in one clean language. They mix Hindi with Marathi, Telugu with English, often in the same sentence. BulBul v3 is specifically trained for this, it handles code-mixed speech reliably, which none of the global alternatives do well out of the box.
For a general knowledge base, standard retrieval works fine. For recipes, there is a specific problem, the same dish can exist in multiple versions across documents, different quantities, different batch sizes, different preparation notes. A basic retrieval system can return the right recipe but for the wrong quantity and may cause hallucination problems.
To solve this specific issue, we used GraphRAG, which maps relationships between entities across all documents rather than just matching keywords. When a staff member asks about a recipe at a specific quantity, the system understands the context and returns the right version. Across hundreds of recipes, this is what makes the answers reliable enough to act on.
Looking ahead
As the chain grows, the system grows with it. New documents added by any branch become available across the whole network. The more the business documents its operations, the more useful the assistant becomes which creates a natural incentive to keep SOPs updated.
Operational intelligence at scale is not something you can retrofit. For any chain planning rapid expansion in India, it is worth building into the foundation early. Our team is now extending the same for various other use cases within the enterprise.
Is your business losing knowledge every time someone walks out the door?
Connect with us to explore how Voice AI can make your institutional knowledge accessible to every employee, in their language, the moment they need it.
