01 — Overview and goals

Duration: 8 min Prerequisites: none. Some Python helps; Course 2 is a plus but not required.

Why this project

Your credit-card transactions are some of the most informative personal data you own. A bank CSV export already contains:

• the date of every purchase,
• the amount (debit or credit),
• the merchant (“Coastal Spoon Cafe”, “Harbor Noodle House”, etc.),
• a category (“Dining”, “Groceries”, “Housing”…).

Banks rarely offer interesting dashboards. LLMs, on the other hand, are great at understanding plain-English questions — but they can’t compute. The right combination:

LLM to understand the question → SQL to compute the answer → LLM to phrase it.

That’s the central idea of this project: an assistant that uses a tool. The LLM doesn’t make up the numbers; it writes SQL, has it executed, reads the result.

Why fully local

You care about	This project gives you
Privacy	Data never leaves your machine
Cost	Zero — no API key, no per-question billing
Offline	Works on a plane, in a cafe, anywhere
Hackability	Swap the LLM with one ollama pull

The trade-off: latency. A question on CPU takes ~5–15 seconds with llama3.1:8b. If you want sub-second answers and don’t mind the cloud, see Project 2.

Architecture in one diagram

   bank CSV (data1-anonymized.csv, 734 rows)
            │
            ▼
   ┌─────────────────────────┐
   │ csv-llm-shared/         │   ← shared Python module (Project 3)
   │  ingest.py              │     • reads CSV
   │  normalize.py           │     • signs amounts, parses dates
   │  db.py                  │     • SQLite + SELECT-only
   │  categorize.py          │     • rules + LLM fallback
   │  dashboards.py          │     • Plotly
   └────────────┬────────────┘
                │
                ▼
       csv-llm-ollama/
         port 8504
         llama3.1:8b
         app.py + llm.py

The Streamlit app has three tabs: Dashboards, Transactions, Chat. The Chat tab exposes one tool to the LLM: query_sql(sql) — read-only on SQLite.

The 4 canonical questions

By the end, your app must answer these 4 questions without hard-coding anything:

1. How much did I spend on groceries last month?
2. What is my biggest spending category?
3. How many times did I visit Coastal Spoon Cafe this year and how much did I spend?
4. List my top 5 merchants.

We’ll verify all four against the real data in chapter 09, with the generated SQL and the final answer.

What you won’t have to do

• Connect to your bank’s API: too much variability, we use a clean CSV export.
• Build your own LLM: we use llama3.1:8b, free and tool-call capable.
• Reinvent a UI: Streamlit does it in under 200 lines.

Takeaways

• An LLM alone doesn’t compute; it writes SQL and reads the result.
• The app exposes one tool: query_sql(sql) read-only on SQLite.
• 100% local: no API key, no cloud, no recurring cost.
• In chapter 02 we’ll look at the demo CSV that ships with the repo.

Next: The demo CSV →