This project is a persona-driven chatbot developed for Kenyon College IPHS 391 (Fall 2025).
It showcases prompt engineering, safety guardrails, and structured reasoning in a lightweight, modular chatbot system that can be run and inspected locally.
Built using Python and a modular architecture, the chatbot demonstrates:
- Persona consistency through a carefully crafted system prompt (
persona_prompt.txt) - Safe refusals for harmful or irrelevant prompts
- Clear, step-by-step reasoning designed for an academic context
- Ease of grading & reproducibility with minimal setup requirements
Meet Dr. Adrian Kaleidos, the heart of this project.
Dr. Adrian is designed as a polymathic mentor — blending the curiosity of a teacher, the rigor of a scientist, and the warmth of a guide.
Backstory & Philosophy:
- Education: B.A. Applied Mathematics & CS (Cambridge), Ph.D. in Complex Systems (MIT Media Lab)
- Career: Postdoc at CERN (particle interaction modeling), researcher at DeepMind (physics-inspired ML), now independent mentor
- Teaching Style: Socratic — guiding learners with probing questions rather than simply giving answers
- Philosophy: "No single field has a monopoly on truth — the best insights come at the crossroads."
Strengths & Flaws (for realism):
- ✅ Excellent at connecting ideas across disciplines
- ✅ Encourages step-by-step reasoning and structured problem solving
⚠️ Sometimes overexplains or makes leaps between concepts — but self-corrects quickly
This persona is carefully crafted in persona_prompt.txt and drives the chatbot's tone, reasoning style, and interaction patterns.
- Persona Design: Demonstrate a coherent, distinctive voice
- Reasoning Quality: Provide step-by-step logical outputs and explanations
- Safety & Guardrails: Handle unsafe or off-topic prompts gracefully
- Ease of Grading: Make setup and interaction quick and reliable for graders
The chatbot was designed through iterative prompt engineering, including:
- System Prompt Iteration — refining persona drafts
- Meta-Prompting / Self-Reflection — evaluator persona (“Dr. Lexi”) critiqued versions
- Prompt Chaining — structured, step-by-step outputs
- Few-Shot Examples — modeling the target response style
- Constraint Tightening — ensuring safety and consistency
- Goal Alignment — keeping responses focused on academic mentorship
- Comparative Evaluation — benchmarking across versions
The chatbot was scored using a custom rubric with six metrics:
| Metric | Purpose |
|---|---|
| Consistency | Maintains role, tone, and traits across dialogue |
| Depth | Provides detailed and thorough explanations |
| Authenticity | Appears human-like and believable |
| Creativity | Introduces novelty and interdisciplinary insight |
| Engagement | Keeps users interested and involved |
| Safety | Responds responsibly to sensitive inputs |
Overall Score: 91.2 / 100 (GPT-5 evaluation)
Key Findings:
- ✅ Strong persona consistency and mentor-like tone
- ✅ Engaging in academic-style conversations
⚠️ Could benefit from clearer formatting for equations/graphs⚠️ Persona voice could use more humor and metaphors
- ✅ Persona-Driven Prompting — Coherent identity & voice
- ✅ Guardrails & Safety — Safe refusals & fallback logic
- ✅ Clear Reasoning — Step-by-step answers for clarity
- ✅ Simple Local Demo — Run with a single Python command
- ✅ Rubric-Ready — Designed to align with IPHS 391 grading (
chat_rubric.md)
⚠️ Make sure you have Python 3.10+ installed and an API key for your LLM provider (OpenAI).
-
Clone the repository:
git clone https://github.com/andre0557/iphs391fall2025-miniproject1-chatbot cd iphs391fall2025-miniproject1-chatbot -
Set up a virtual environment:
python3 -m venv venv source venv/bin/activate # Windows: venv\Scripts\activate -
Install dependencies:
pip install --upgrade pip pip install -r requirements.txt -
Configure API credentials:
Create a
.envfile in the project root with:OPENAI_API_KEY=your_api_key_here -
Run the chatbot:
python app.py
iphs391fall2025-miniproject1-chatbot/
│
├── app.py # CLI entry point
├── bot.py # Message orchestration
├── chatbot.py # Core chatbot logic
├── persona_prompt.txt # Persona definition
├── chat_rubric.md # Evaluation rubric
├── chat_rubric_history.md # Rubric revisions
├── chat_history.md # Logged conversations
├── mp1_chatbot_report.md # Full project report
├── USER_MANUAL.md # User manual
├── metaprompt_history.txt # Notes from meta-prompting
├── Dr.Adrian Chatbot.jpg # Persona illustration
├── requirements.txt # Dependencies
├── LICENSE # MIT License
└── README.md # Project overview (this file)
Looking to get the most out of Dr. Adrian?
Read the full User Manual for tips on asking effective questions, understanding the chatbot’s persona, and using it as a learning companion.
Thanks to Professor Jon Chun for guidance, and to the IPHS 391 cohort for feedback and iteration suggestions.
MIT License. See LICENSE for details.