How much Python do you need for AI engineering? A tiered answer

Tier 1: essential from day one

You will use these in almost every task.

Tier 2: needed for production work

These are what most interview take-home tasks and code reviews check.

Tier 3: can wait

Useful later, rarely the blocker for an entry or mid-level AI engineering role:

• metaclasses, descriptors and advanced decorators beyond the basics,
• C extensions and performance tuning of Python itself,
• writing your own async framework,
• deep NumPy broadcasting tricks (unless you work on model training).

Learn them when a real problem needs them.

A five-minute self-test

Read this function and predict the output before running it. It parses a model's structured output — a very common AI-engineering task — and it contains the Tier 1 and 2 ideas in miniature.

import json
from dataclasses import dataclass

@dataclass
class Invoice:
    number: str
    total: float

def parse_invoice(raw: str) -> Invoice | None:
    """Return an Invoice, or None if the model output is unusable."""
    try:
        data = json.loads(raw)
    except json.JSONDecodeError:
        return None
    if not isinstance(data, dict):
        return None
    number, total = data.get("number"), data.get("total")
    if not isinstance(number, str) or not number.strip():
        return None
    try:
        total = float(total)
    except (TypeError, ValueError):
        return None
    if total < 0:
        return None
    return Invoice(number.strip(), round(total, 2))

cases = [
    '{"number": "INV-104", "total": "18450.50"}',   # total as string
    '{"number": "INV-105", "total": -20}',          # negative
    'Sure! Here is the JSON: {"number": "INV-106"}',# prose around JSON
    '["INV-107", 900]',                             # wrong shape
    '{"number": " INV-108 ", "total": 999.999}',    # whitespace, rounding
]
results = [parse_invoice(c) for c in cases]
for c, r in zip(cases, results):
    print(r)

assert results[0] == Invoice("INV-104", 18450.5)
assert results[1] is None and results[2] is None and results[3] is None
assert results[4] == Invoice("INV-108", 1000.0)

If you predicted all five correctly, and you could explain why case 3 fails and how you might handle it (for example, by asking the model for strict JSON output, or extracting the JSON object before parsing), your Tier 1 is solid. If you would also write tests for these cases and log the rejected ones, you are working at Tier 2.

What interviewers check

In AI engineering take-home tasks and live coding, reviewers commonly look for:

1. 1Handling bad model output without crashing, as above.
2. 2Timeouts and retries around API calls, with a limit.
3. 3Secrets out of code.
4. 4Tests for the parts that are deterministic.
5. 5Readable structure — small functions with clear names.

Clever one-liners rarely help; clear, defensive code does.

A learning order that works

If you are starting from basic Python:

1. 1Weeks 1–2: Tier 1, using small scripts that read files and JSON.
2. 2Weeks 3–4: call a model API; add error handling, timeouts and logging.
3. 3Weeks 5–6: dataclasses or Pydantic for output validation; write pytest tests.
4. 4Weeks 7–8: asyncio for concurrent calls with a concurrency limit; package the project.

By the end you have one project that demonstrates every Tier 2 skill, which is better evidence than several tutorials.

Is TypeScript needed too?

Some AI engineering roles, especially those building web products or agent tool servers, also use TypeScript. See TypeScript for AI engineers to judge whether it matters for your target roles.

The Generative & Agentic AI add-on assumes working Python and builds Tier 2 habits through its projects. If you are coming from data science, what to relearn when moving to AI engineering explains which of your habits to keep.