Skip to content

Portfolio Chatbot

1 · Motivation 🤔

I wanted something smarter than a static “About Me” page—
a way for visitors to ask me anything about my background, projects, or skills,
and for me to know when someone’s genuinely interested.

But I didn’t want to:

  • hard-code a brittle FAQ
  • pay for a full vector database
  • miss potential leads while I’m offline

So I built a lean, 150-line Python assistant that:

  • loads my résumé from a simple JSON file (no DB)
  • routes queries to the right section (projects, skills, etc.)
  • calls GPT-4o-mini using OpenAI function calling
  • alerts me via Pushover when someone submits a question it can’t answer, or shares their email

The result is a minimal, self-contained chatbot for my personal site that runs without a database and only needs an OpenAI key and a Pushover token to function.

2 · Architecture ⚙

Chatbot architecture

Key decisions

Problem Why not X Solution
Retrieval for a 5 KB résumé Vector DB overkill Regex router + load JSON in-memory
Unknown questions Manual inbox checking record_unknown_question tool → push
Lead capture Contact form spam GPT asks for email + one-liner, then record_user_details

3 · Knowledge Base Structure 📚

data/chunks.json
[
  { "id": "summary",
    "section": "summary",
    "text": "Bangalore-based CS undergrad focusing on applied ML…" },

  { "id": "exp_videoverse",
    "section": "experience",
    "text": "ML Engineer Intern — VideoVerse (Aug–Nov 2024)…" },

  { "id": "prj_fire_detection",
    "section": "project",
    "repo": "https://github.com/NeuralNoble/fire-detection",
    "text": "Drone fire-detection system using YOLOv8-Nano…" },

  { "id": "skills_core",
    "section": "skills",
    "text": "Python · PyTorch · TensorFlow · AWS · Docker…" }
]

Note

Keep each chunk ≤ 350 tokens so three of them + the prompt fit comfortably under GPT-4o-mini’s context window.

4 · Core Logic 💻

Router & context builder 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
INTENT_RE = {
    r"\b(project|github|built)\b": "project",
    r"\b(experience|job|intern)\b": "experience",
    r"\b(skill|tech|tool)\b": "skills",
    r"\b(educat|degree|study)\b": "education",
}

def route(query: str):
    for pattern, section in INTENT_RE.items():
        if re.search(pattern, query.lower()):
            return section
    return None  # fallback

def build_context(query: str):
    section = route(query)
    parts = [c["text"] for c in CHUNKS if c["section"] == section] \
            if section else [c["text"] for c in CHUNKS]
    return "\n\n".join(parts)
Pushover helpers
def push(msg):  ...
def record_user_details(email, name="n/a", notes=""): push(...)
def record_unknown_question(question): push(...)
LLM call
messages = [
    {"role":"system", "content": persona},
    {"role":"system", "content": f"Context:\n{ctx}"},
    {"role":"user",   "content": user_msg}
]
response = openai.chat.completions.create(
    model="gpt-4o-mini",
    messages=messages,
    tools=TOOLS,
    max_tokens=256
)

5 · Running Locally 🚀

git clone https://github.com/NeuralNoble/portfolio-bot
cd portfolio-bot
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# add OPENAI_API_KEY + Pushover keys to .env
python chat.py

Open http://localhost:7860 and start a conversation.

6 · Things I’d Improve Next 💡

  • Swap regex routing for a MiniLM classifier to catch synonyms (“stack”, “tech”).
  • Add a Whisper endpoint for voice queries.
  • Cache answers in SQLite to cut token cost for repeated FAQs.

Questions, suggestions, or want to fork it? Ping me on LinkedIn or open an issue on GitHub.