Authors: Tejaswini K, Sameer Mokhasi, Laksh G, Rajat Y
Abstract: For founders to fundraise, they need to find investors that fit the same stage, sector, and geography as their company, contact them via an email, and track responses in multiple disconnected tools. Advibe is a prototype that unifies all these tasks in a unified workflow. The system takes a pitch deck or company description, creates a company profile, finds suitable investors in a catalog, scores them using an explainable score, drafts emails to be sent to the investors, and classifies the responses received. The Advibe system has a React and Vite client with a FastAPI backend, using PostgreSQL database on Supabase and email delivery through Resend. All language models are used to create company profiles, score investors and give an explanation, draft emails, and classify responses. The main model used is Llama 3.3 70B versatile which is served by Groq, and the fallback model is DeepSeek-R1 served through OpenRouter while the fallback for keyword-based use cases for company profile and response classification is also used. Ranking involves two stages: the rule score R, based on stage, sector, and geography matches, plus recency, ranks up to 50 candidates, who in turn get ranked by language model score L, with final score S = 0.40R + 0.60L. In our analysis of this score in closed form we enumerate all seven possible rule scores and compute break-even language model score for each of them. The investor catalogue contains 969 investor entries and 1,242 contact entries, of which the first sheet of the workbook supplied for analysis lists 114 investors, 84.2% of them residing in India. Intake, matching, drafting, approval gate for sending, and replies handling activities are related through 17 backend routes modules and 24 client components. For 20 test profiles, LLM-judged rating puts the rule stage significantly above random ranking (precision@10 = 0.91 vs 0.42). Human judgements and evaluation of LLM stage require future work.
DOI:
