OfBusiness built Nexizo on a fine-tuned Llama, not an LLM wrapper (interview)

Gurugram B2B ecommerce and financing group selling raw materials such as steel and chemicals to SMEs, which also runs fintech arm Oxyzo.

Fine-tuning a small open model on OfBusiness's own tender and price data, sold as a separate SaaS, beats a wrapper over someone else's LLM.

No new capital; Jain said OfBusiness did not hire new talent for Nexizo and upskilled existing staff.

In an interview with Inc42, Jain says OfBusiness already used machine-learning models to show government tenders, and after ChatGPT launched a chat assistant for raw-material prices and commodity news, updating 10,000 commodity prices in real time.

Jain says the assistant was only a layer over an existing LLM, so the company fine-tuned Llama to build its own small language models for tender classification and information, starting training toward the end of 2023 and running a beta for almost a year before launching Nexizo.

Nexizo lets SME and contractor users search tenders by natural language and generates leads for NBFCs and raw-material suppliers. Jain says it has 1M registered users, about 400K monthly actives and 20 to 25 enterprise clients (self-reported in the interview), and the company is still experimenting with monetisation.

A conversational layer over a third-party LLM, Jain says, offered no differentiation, so OfBusiness moved to its own fine-tuned models.

Infrastructure tenders account for about 30% of OfBusiness's revenue, and its tender data gave it a proprietary training set.

B2B orders of INR 10 Lakh to INR 20 Lakh keep human sales teams involved, so Nexizo is positioned as a lead generator.

An AI product stays defensible when it is fed by proprietary data you already own; a thin layer over someone else's model is not a moat.

As of 30 March 2025, Nexizo was live with about 400K monthly active users per Jain, BidAssist was expected to integrate into it, and OfBusiness was still experimenting with a model that sells leads instead of charging for tender data.