Published Paper Details:

Pre-seed startup evaluation is a messy problem. There is no funding history to analyse, no revenue to model, and the founder's pitch deck may be the only artefact that exists. Venture capitalists still need to assess market size, map competitors, stress-test the business model, and flag risks--but at this stage, much of that work amounts to educated guesswork. Large language models can generate business analyses that read well, yet their outputs suffer from two recurring failures: structural inconsistency (reports that cover financials one time and skip them the next) and hallucinated data (invented market figures, nonexistent competitors). This paper introduces ValidatorAI, a system that addresses both problems by constraining LLM output through a seven-dimension JSON schema derived from standard VC due-diligence checklists and grounding generation in live web search via the Gemini API. We evaluated three configurations against 50 startup concepts spanning ten sectors, scored by four independent raters on factual accuracy (FA), analytical completeness (AC), and decision quality (DQ) on a 1-5 Likert scale (Cohen's ? = 0.76). A 20-concept held-out subset was also run on OpenAI's GPT-4o. The full pipeline (C3) achieved FA = 3.9, AC = 4.1, DQ = 3.7--a 41% accuracy gain and 28% completeness gain over the unconstrained baseline. GPT-4o under the same schema matched on completeness (AC = 4.0) but scored lower on accuracy (FA = 3.2), lacking native search grounding. Paired t-tests confirmed significance for all C1-to-C3 comparisons (p < 0.01). The two mechanisms fix different things: schema enforcement drives completeness; search grounding drives accuracy. Neither alone was enough.