Project 2 — Card-spending analysis with OpenAI (cloud)

Audience: students who finished Project 1 or want the fastest version of the spending analyzer. ~95% of the code is identical; only llm.py differs.

Level · IntermediateDuration · ~1.5 hoursChapters · 7Stack · Streamlit + Python + SQLite + OpenAI API + Plotly

Why a cloud variant

You want	You pick
Privacy, no API key, no recurring cost	Project 1 — Ollama
Sub-second answers, max accuracy, OK with cloud	This project (OpenAI)

The OpenAI variant is built specifically to answer the question: “How would the same app look if I replaced ollama.chat() with openai.chat.completions.create()?”

Spoiler: the diff is ~40 lines. Everything else (ingestion, normalization, SQLite, dashboards, the query_sql tool spec, the system prompt) is byte-for-byte identical.

Architecture (same as Project 1, different LLM)

   bank CSV (data1-anonymized.csv, 734 rows)
            │
            ▼
   ┌─────────────────────────┐
   │ csv-llm-shared/         │   ← shared Python module (Project 3)
   └────────────┬────────────┘
                │
                ▼
       csv-llm-openai/
         port 8505               ← (Project 1 was 8504)
         gpt-4o-mini
         app.py + llm.py        ← only llm.py differs from Project 1

The two apps can even run side-by-side on the same machine: csv-llm-ollama on 127.0.0.1:8504, csv-llm-openai on 127.0.0.1:8505. Same SQLite schema, separate .cache/ databases.

What you’ll have at the end

• A Streamlit app on http://127.0.0.1:8505 that:

  • loads csv-llm-openai/data/data1-anonymized.csv;
  • shows the same 4 dashboards as Project 1;
  • answers your 4 canonical questions in ~2 seconds instead of ~10.

• A clear understanding of:

  • the 4 ways openai.chat.completions.create() differs from ollama.chat(),
  • the cost per question (spoiler: under $0.001),
  • the privacy trade-off and how to mitigate it,
  • the .env + .gitignore pattern to never leak an API key.

The 7 chapters

#	Chapter	Duration	Type
01	Overview and goals	8 min	Theory
02	Setting up the environment + API key	10 min	Hands-on
03	llm.py side-by-side: Ollama vs OpenAI	15 min	Code
04	The OpenAI agent loop with SQL tool	20 min	Code
05	Running the app step-by-step	12 min	Guided
06	Cost, latency, accuracy comparison	10 min	Analysis
07	API key security and going further	10 min	Theory

Prerequisites checklist

Before starting:

• ✅ You’ve read at least chapters 04–08 of Project 1, or you’re comfortable with ingest.py / db.safe_query() / the SQL_TOOL spec.
• ✅ You have an OpenAI account with at least $1 of credit (one cent is enough for the whole course; $1 is the minimum payment).
• ✅ Python 3.9+, pip, and the csv-llm-shared/ repo already cloned.

How to read this course

• Linear: 01 → 07. About 1.5 hours.
• Already finished Project 1: skip 01, jump to 03. About 1 hour.
• Just run the app: 02 → 05. 25 min.

← Coming from Project 1 — Ollama? You’ll feel right at home. → Want to deep-dive the shared library powering both apps? Head to Project 3 — csv-llm-shared.

Start with chapter 01 →