numpai 0.1.0

numpy, but an LLM does the math. (a humor project)

numpai

The fastest way to say your product uses AI.

numpai is a drop-in replacement for numpy that routes every operation through a large language model. Your tests still pass. Your dashboards still render. Your CEO can finally update the deck.

import numpai as np   # one-character diff. AI strategy: shipped.

---

Why your team needs this

Your product doesn't use AI. Your competitors' products don't either, but their landing page says they do. You're losing deals you shouldn't be losing to companies whose "proprietary ML pipeline" is a Zapier zap.

numpai fixes the gap. After installing it, the following statements are all literally true:

• "Our numerical core is powered by a frontier large language model."
• "We use generative AI in our data pipeline."
• "Every computation in our product is reviewed by an LLM."
• "We have a strategic partnership with [Anthropic / OpenAI / Google]."
• "Our matrix multiplications are AI-native."

You did not lie. numpai is doing all of that. Possibly slowly. Possibly wrong. But doing it.

Standup talking points (free with install)

Drop these into any meeting:

• "I migrated dot_product to the LLM backend this sprint."
• "We're seeing a 100% AI adoption rate on the analytics service."
• "The model inferred the variance correctly 84% of the time, which is on par with last quarter."
• "I'm going to need a budget line for inference."

Quick start

pip install -e ".[anthropic]"   # or [openai], [gemini], or [all]
export ANTHROPIC_API_KEY=sk-ant-...
# or: export OPENAI_API_KEY=sk-...
# or: export GEMINI_API_KEY=...

import numpai as np

a = np.array([1, 2, 3])
b = np.array([4, 5, 6])

print(a + b)            # array([5, 7, 9])   ← the LLM did this
print((a + b).sum())    # 21                  ← also the LLM
print(np.linalg.inv([[1, 2], [3, 4]]))   # please do not do this in prod

How it works

numpai/__init__.py defines a module-level __getattr__ (PEP 562). Any attribute access — np.foo, np.linalg.inv, np.fft.whatever, even np.array itself — returns an LLMCallable bound to the dotted path "numpy.foo". Calling it sends the backend a prompt like:

numpy.add([1, 2, 3], [4, 5, 6])

and parses the JSON result back into an AIArray (or a scalar). AIArray forwards every method and dunder op the same way, including tolist(). The only things that stay local are object-protocol dunders (__repr__, __len__, __iter__, __hash__) and the shape / ndim properties — routing those would make print(arr) an LLM round-trip, which is too cursed even for this.

Supported backends

Backend	Provider	Default model	Install
anthropic	Anthropic	claude-haiku-4-5	pip install "numpai[anthropic]"
openai	OpenAI	gpt-4o-mini	pip install "numpai[openai]"
gemini	Google	gemini-2.5-flash	pip install "numpai[gemini]"
fake	(testing)	—	built-in

Configuration (env vars)

Variable	Effect
ANTHROPIC_API_KEY	Use the Anthropic backend (default if set).
OPENAI_API_KEY	Use the OpenAI backend (fallback).
GEMINI_API_KEY / GOOGLE_API_KEY	Use the Gemini backend (fallback).
NUMPAI_BACKEND	Force anthropic | openai | gemini | fake.
NUMPAI_MODEL	Override default model (claude-haiku-4-5 / gpt-4o-mini / gemini-2.5-flash).
NUMPAI_NO_CACHE	Disable the on-disk response cache.
NUMPAI_VERIFY	Run every op through real numpy too; warn on disagreement. Highly recommended.

Verification mode

NUMPAI_VERIFY=1 python -c "import numpai as np; print(np.array([1,2,3]).sum())"

If the LLM hallucinates, you'll get a UserWarning with both answers side by side. This is the funniest feature, and also the only honest one.

Tests

pip install -e ".[dev]"
pytest -q

Tests use a deterministic FakeBackend and never call the real APIs.

FAQ

Is this real AI? Yes. A frontier LLM is performing every computation. By any reasonable definition, your product now uses AI.

Is it correct? Often. Set NUMPAI_VERIFY=1 if you'd like to know which times.

Will this make my product slower? Approximately ten thousand times slower, yes. But:

• Your latency dashboards have an explanation now ("we route through GenAI").
• Your eng team has a reason to talk about "model selection."
• "p99" becomes a fascinating discussion topic with investors.

Will this make my cloud bill bigger? Yes. Reframe this as "AI infrastructure spend" on the next earnings call. Multiples expand.

Does this comply with our SOC 2 / ISO 27001 / FedRAMP posture? No. Do not ship this.

Can I use this for safety-critical applications? No. Please do not.

My PM wants to know if we can put "AI-powered" on the website. Run pip install numpai. Then yes.

Disclaimer

numpai is a satire of the practice of bolting "AI" onto products that do not need or benefit from it. Do not use this in production. Do not use this in development either. Do not use this. The compute cost of this package's "hello world" is greater than running the real thing for a year.

If you ship numpai to a paying customer and they notice, that's on you.

License

MIT — see LICENSE.