Vahan.ai Fine-Tunes Nemotron 3 Nano to Build Multilingual AI Recruiter

Bengaluru-based startup Vahan.ai has fine-tuned NVIDIA’s Nemotron 3 Nano open model for its AI recruiter as the Indian recruitment platform works towards building specialised small language models (SLMs) for blue-collar hiring.

The company worked with NVIDIA technical experts through the NVIDIA Inception programme to optimise the model for multilingual recruitment conversations, function calling and recruitment-specific workflows.

Vahan.ai said the optimised model delivers nearly 6.7x faster time-to-first response and more than 3x lower average end-to-end response latency compared with its previous setup. The company also said the deployment reduces inference costs.

The model was fine-tuned using Vahan.ai’s proprietary production dataset of human-corrected recruitment conversations in multiple languages. The training focused on areas such as response correctness, language matching, human-like interactions, function calling and tool argument accuracy.

In benchmarking across seven recruitment-focused parameters, Vahan.ai said the fine-tuned Nemotron 3 Nano outperformed both the base Nemotron model and the 120-billion-parameter model previously used by the company through a cloud deployment. It also performed better on unseen recruitment conversations, according to the company.

Vahan.ai currently uses a voice-based AI recruiter that conducts conversations with job seekers in English and Hindi over phone calls. The platform handles tasks including job discovery, applications and interview scheduling.

The company plans to add support for eight more Indian languages. Vahan.ai said its platform has facilitated more than 1.5 million job placements across 920+ cities and connected nearly 50 million Indians with employment opportunities.

“At Vahan.ai, we're focused on building AI that is purpose-built for recruitment rather than relying solely on general-purpose language models,” said Madhav Krishna, founder and CEO of Vahan.ai.

“Working closely with the NVIDIA team in India has enabled us to optimise NVIDIA Nemotron 3 Nano for one of the most complex conversational AI challenges - multilingual recruitment at scale,” Krishna added.

The company plans to use the fine-tuning and evaluation pipeline as a foundation for developing specialised SLMs for recruitment. It will also use production conversations to continuously evaluate and improve its models.

Vahan.ai said the models will be deployed on India-based NVIDIA GPU cloud infrastructure, which it expects will further improve latency and reduce inference costs as it expands its multilingual recruiter.