Multilingual NLU systems with Saarthi.ai
Saarthi.ai runs conversational AI bots for enterprises in India, and those bots have to understand people in the languages they actually speak. We worked with them on the multilingual NLU and NER underneath, which went into production across more than 15 client deployments.
The difficulty is that no single model covers it. Intent and entity extraction has to hold across Hindi, Tamil, Bengali, and the way people really mix and vary them. On top of that, every enterprise brings its own vocabulary of product names, locations and account terms that a general model has never seen.
Why One Model Was Never Going to Be Enough
There were two constraints pulling in different directions. Coverage needed to be broad, across languages and across each client’s private vocabulary. But the bots run inside live conversations, so extraction also had to be fast enough for natural turn-taking.
A single large model can chase coverage, but it is slow and still misses a client’s niche terms. So we split the work by what each method is good at.
Rules for What Must Never Be Missed
Some entities have known shapes: account numbers, dates, specific product codes. Missing one of those breaks the conversation, so we handled them with rule-based NER built from regex and pattern dictionaries. Rules do not guess, which is exactly what you want for these.
Deep learning covers the rest: intent, and the language variation that rules cannot reach. We also refactored the Python pipelines for NER data creation and the multilingual NLU code paths. That meant a new client’s vocabulary could go into the pattern layer without anyone touching model code.
Making the Models Small Enough to Keep Up
Latency work centred on knowledge distillation. A large model teaches a smaller one to mimic its answers, so you keep most of the capability at a fraction of the cost. We used it to move capability into models that meet production response budgets, with transfer learning shaped by each deployment’s data.
For the enterprise chatbot cases that need it, LLM integration sits on top of the pipeline rather than replacing it.
What It Meant in Production
- Multilingual NLU and NER deployed across 15+ client production systems.
- Better entity extraction and intent coverage from rule-based regex patterns and custom dictionaries.
- Lower production model latency through knowledge distillation.
- Support for large-scale LLM integration in enterprise chatbot deployments.
The work was built in Python with deep learning, knowledge distillation, transfer learning, NER, NLU, and LLM integration.