Week 1 — Introduction to AI Research

What is Research?

• Research vs. coursework: the difference
• What is a research question? What is a contribution?
• What is a publication? What is a conference?
• What is a PhD and why would someone want one?

What is ComplianceGPT?

• The problem: healthcare compliance is expensive and error-prone
• The solution: AI that reads scenarios and gives verdicts
• Live demo: run the pipeline, see a PERMITTED / DENIED result
• The pipeline diagram: NL → LLM → Engine → Verdict

Variables & Types

name = "hospital"
count = 137
is_permitted = True

Lists & Dicts

models = ["gemma3", "llama3"]
result = {
  "verdict": "PERMITTED"
}

Functions

def accuracy(df):
    correct = (df["v"] ==
               df["gt"]).sum()
    return correct / len(df)

The AI Stack

• AI — the broad field
• Machine Learning — learning from data
• Deep Learning — neural networks
• Generative AI — creating text/images/code
• LLMs — what we use in ComplianceGPT

What LLMs Can and Cannot Do

• Can: read long text, extract structure, reason step-by-step
• Cannot: guarantee factual accuracy (hallucination)
• Cannot: reason perfectly about legal edge cases
• Cannot: reliably follow complex instructions (sometimes)