Romanian-Language Conversational Sales Agent (SalesGPT adaptation)
Overview
A build conversational sales agent adapted from the open-source SalesGPT framework and localized into Romanian. It role-plays a sales representative for a specific business, moving through structured conversation stages (introduction, qualification, value proposition, objection handling, close).
Why It Exists
To test whether an autonomous, stage-aware sales agent can hold a natural Romanian-language conversation, qualify a prospect, and pitch a real business, exploring LLM-driven outbound/inbound sales automation for a local-market context.
What We Built
A SalesGPT-based agent (ro.py) configured with custom Romanian prompts, a defined salesperson persona and company profile (a Bucharest family restaurant), and a product-catalog knowledge base under examples/. The agent uses LiteLLM to stay provider-agnostic across 50+ LLMs, supports tool use, and tracks conversation stage explicitly before each turn. Translated prompt scaffolding (ro.txt) handles the full sales-stage playbook in Romanian.
Technologies & Approach
Python on top of the SalesGPT/LangChain stack, with LiteLLM abstracting the model layer (GPT-3.5/4 in this run). The key custom work is the Romanian localization of the entire conversation-stage prompt system and the business-specific persona/catalog grounding.
Outcome / Impact
A working demo proving that a stage-aware, tool-using sales agent can be localized and grounded in a real business. Validates conversational-agent prompt engineering and non-English LLM deployment for the studio.
Capabilities Demonstrated
- Building stage-aware conversational sales/outreach agents
- Localizing LLM agents (Romanian-language prompt engineering)
- Provider-agnostic LLM integration via LiteLLM
- Grounding agents in business-specific catalogs and personas