+ $ speedcraft optimize slow_script.py
+ Profiling slow_script.py ...
+ Bottleneck: is_prime_count — 0.0019s cumulative
+ Requesting Cython translation from LLM...
+ Compiling generated Cython code...
+ Running differential test (300 random trials)...
+ ✓ PASSED 300/300 trials — correctness verified.
+ ⚡ Speedup: 19.6xPython developers write slow loops. Making them fast usually means learning Cython's type system and hand-porting code to C — a skill most Python devs never get around to.
SpeedCraft-AI automates the whole pipeline: it profiles your script, finds the actual bottleneck functions, asks an LLM to translate them into typed Cython, compiles them with gcc, and — before any of it touches your code — runs 200+ randomized differential tests comparing the original and compiled output. If they don't match exactly, the conversion is thrown away and your original Python stays untouched.
No manual C. No blind trust in AI-generated code either.
This is the part that makes the tool trustworthy instead of just fast:
| Stage | What it does |
|---|---|
| Eligibility gate (AST-based) | Rejects functions with try/except, generators, closures, dicts/strings, or I/O calls before ever calling the LLM |
| Differential testing | Runs the original and compiled function on identical randomized inputs, hundreds of times, comparing outputs with float-tolerant equality |
| Automatic fallback | Any mismatch → conversion is discarded, original Python function is kept, reason is printed |
Real example from development: an early Cython template used
int(num ** 0.5)for a primality check. Floating-pointpow()precision occasionally returned4.999999...instead of5.0, silently misclassifying25as prime. The differential tester caught it in 3 trials. This is exactly the class of bug that "looks right" testing misses.
cProfile → Eligibility (AST) → LLM → Cython → gcc compile → Differential Test (200x)
│
┌───────────┴───────────┐
PASS FAIL
│ │
persist .so + write shim keep original Python
✅ Handles: top-level, pure numeric, loop-heavy functions — prime counting, simulations, custom math, nested loops.
❌ Does not handle (yet): classes, generators, closures, string/dict-heavy code, I/O. These are rejected with a clear reason instead of silently mishandled.
This is not a Numba/Codon replacement — if your code is array-heavy numeric math, use those, they're mature and don't need an LLM in the loop. SpeedCraft-AI targets the gap: custom scalar/loop logic that doesn't fit neatly into @jit.
git clone https://github.com/fay-saal/speedcraft-ai.git
cd speedcraft-ai
pip install -e .
export ANTHROPIC_API_KEY=your_key_hereRequires a C compiler (gcc on Linux/Mac, MSVC Build Tools on Windows) — the same requirement Cython itself has.
# profile + optimize (default: top 5 functions, 200 test trials)
speedcraft optimize my_script.py
# preview what would be converted, no API calls, no cost
speedcraft optimize my_script.py --dry-run
# tune rigor
speedcraft optimize my_script.py --top 3 --trials 500Then swap your import:
# before
from my_script import slow_function
# after
from my_script_accelerated import slow_function
v0.1, early release. Core pipeline (profiling, eligibility checking, compilation, differential testing) is tested end-to-end. LLM conversion prompt is functional but still being refined based on real-world feedback — issues/PRs welcome.
MIT © 2026 Faysal (DieBack Theatre)
- Repository: github.com/fay-saal/speedcraft-ai
- Issues: github.com/fay-saal/speedcraft-ai/issues