Vahan AI fine-tunes 30-billion-parameter Nvidia Nemotron model for blue-collar hiring

Vahan.ai, an AI recruitment platform for blue-collar and gig workers, has fine-tuned Nvidia’s Nemotron 3 Nano model for its voice-based recruiter to make conversations with job seekers faster, cheaper and accurate across Indian languages.

The company worked with Nvidia’s technical team through its Inception programme to customise the 30-billion-parameter model using Vahan’s proprietary recruitment data. The model has been deployed in production and is currently handling about 10% of the company’s traffic, founder and CEO Madhav Krishna told the Economic Times Digital in an interview.

“We have now deployed it in production and we are diverting about 10% of our traffic to it. We are seeing early green shoots for sure. But the scale will come over time,” Krishna said.


Vahan operates a voice-based AI recruiter that speaks to blue-collar job seekers, matches them with jobs and helps them through parts of the hiring process. The company said it has facilitated more than 1.5 million placements across over 920 cities and towns in India.

Making voice AI faster

Vahan said the fine-tuned Nemotron model delivered nearly 6.7 times faster time to first response and more than three times lower average end-to-end latency compared with the system it was previously using.

Krishna said the improvement matters particularly for voice applications, where delays between a user speaking and the AI replying can make conversations feel unnatural.

“There is generally a 500 millisecond to a second long latency, which makes the experience a little unnatural,” he said. “So that will get cut down significantly and it will be a lot more natural for the user.”

The company had previously been using an off-the-shelf 120-billion-parameter language model for the same application, Krishna said.

Vahan said its benchmarks tested areas including response correctness, human-like responses, language matching, function calling and the accuracy with which the AI passes instructions to other tools. The company said the fine-tuned model performed better than both the base Nemotron model and its previous cloud-hosted model.

“So across all the benchmarks that matter for the AI recruiter, we saw 2 to 3x improvement in performance, and we saw about a 7x reduction in time to first response,” Krishna said.

Training AI for Indian recruitment

Vahan used its repository of conversations with blue-collar job seekers to prepare the training data. Krishna said the dataset used for the exercise amounted to around 20,000 or 30,000 hours of calls.

According to Krishna, a key reason for fine-tuning a specialised model was the difficulty general-purpose models have in handling the way Indian users switch languages and use regional variations.

“In Hindi, the word 'haan' (yes) can be said in 20 different ways, depending on where you are from,” he said. “For our use case, we are now able to capture a lot of that nuance.”

Vahan said it will run the model on Nvidia GPU infrastructure through an India-based cloud provider. Krishna said this would also mean that “the data then doesn’t leave the country at all.”

The company is also working with Nvidia on fine-tuning open-source speech models for other parts of its voice stack, while building towards specialised small language models for recruitment.