CohortFilter AI: Accelerator Application Screener

Startup accelerators receive hundreds to thousands of applications every cohort. Program managers spend most of their time on manual triage, reading one by one, with no consistent scoring and no integrated AI tool to help. CohortFilter AI solves this directly. Managers describe their criteria in a short conversation and receive a weighted scoring rubric. They upload a CSV from Typeform or Google Forms, and the system scores every application in parallel, generates a written explanation per score, flags duplicate positioning, and marks unreachable websites. The output is a ranked dashboard with color-coded tiers, per-application reasoning, and one-click export to PDF and CSV. The product runs on AMD MI300X via AMD Developer Cloud, using vLLM on ROCm for parallel batch inference at full 70B model precision without quantization. The backend is FastAPI with async job processing and a clean single-page frontend built for B2B operators. The core logic is domain-agnostic. The same rubric-first framework applies to resume screening, research paper selection, grant applications, and any domain requiring structured decisions across large input sets. Data stays within the operator's environment by design. Built in 72 hours to demonstrate that thoughtful AI-assisted decision-making, on open hardware and an open stack, can replace weeks of manual triage without removing human judgment from the final call.