RPEPTFS - Role-playing Enhanced Pitch Training and Feedback Simulator

About RPEPTFS

RPEPTFS is an AI-powered pitch-practice platform. It lets people rehearse and refine investment pitches through interactive, role-playing sessions with AI investor personas styled after well-known "Dragons' Den" and "Shark Tank" investors, drawn from publicly available information. The same approach can be extended to other personas for non-profit or for-profit fundraising practice.

RPEPTFS is intentionally built as a practical, real-world demonstration of three underlying open-source projects by Hawke Robinson: AILCPH (large-context project helper), SIIMPAF (the self-hosted AI familiar/framework), and DGPUNET (a distributed GPU network). It runs entirely on self-hosted infrastructure with no cloud dependencies.

Status: This is early-stage R&D, not production-ready. It is intended for internal use and open-source collaboration rather than public Internet-facing hosting at this time. Several capabilities described below are functional in alpha; others are in progress.

What It Does

Realistic Investor Personas

A panel of AI investor characters, each with a distinct personality, investment focus, and questioning style, served by self-hosted local LLMs with persona-specific fine-tuned adapters.

Individual & Panel Practice

Practice one-on-one with a single investor or face a full panel of up to five at once, mirroring a real pitch room.

Interactive Chat Sessions

Turn-based back-and-forth conversation with investors, with full conversation history maintained throughout the session, plus pause and resume.

AI-Suggested Responses

Optional follow-up suggestions generated by an assisting agent, to help you consider how to respond during a pitch.

Personality Modeling

Persona behavior is shaped using a Big Five (OCEAN) psychological model combined with RoleMaster-style attributes and investment modifiers, aiming for more rounded, distinctive personas.

Feedback & Progress

Per-investor feedback, interest-level tracking, optional deal offers with amount and equity, decision reasoning, and progress tracking across multiple sessions.

How It Fits Together

RPEPTFS is a layered, OOAD-oriented system intended to stay modular and maintainable:

Web UI (React + TypeScript) for pitcher management, chat, and session history.
Backend API (Java / Spring Boot) coordinating pitchers, sessions, and panel evaluation, backed by PostgreSQL.
Agent Orchestrator (Python / FastAPI) running on GPU servers, coordinating the individual investor agents.
Local LLM Inference (vLLM) serving the persona models, with no cloud dependency.
Distributed GPU via DGPUNET, so multiple simultaneous investor personas (and, over time, animated avatar responses) can be served beyond what a single GPU can hold.

Why DGPUNET matters here: running several simultaneous investor personas (and the planned real-time animated avatars) would overwhelm a single GPU. Spreading the workload across DGPUNET's pooled VRAM is what makes the full multi-persona experience practical.

Privacy & Principles

100% self-hosted, no cloud dependencies, with your data staying on your own infrastructure.
Open source, intended to use OSI-approved licenses.
Personas are built only from publicly available information.
Designed with object-oriented best practices for a clean, maintainable, extensible architecture.

Technology Stack

React + TypeScript Java / Spring Boot Python / FastAPI PostgreSQL vLLM LoRA / QLoRA adapters DGPUNET CUDA 12.x SIIMPAF AILCPH

Built on Hawke's AI Suite

RPEPTFS is one practical application of a larger set of open-source, self-hosted AI projects. Each can be used on its own or together: