ಅಯಸ್ಕಾಂತ · ayaskānta — lodestone; the stone that draws iron
AI-skantha builds small, offline-first models for problems that matter here — privacy, message fraud, bank documents, farm advisory in Indian languages — and Chukki, a desktop app that puts frontier AI behind a prepaid UPI wallet.
Open models, free on Hugging Face · Chukki in private beta
Create bubble_sort.py with a bubble sort and a test block.
Done — created bubble_sort.py with the function, a docstring and a __main__ test. Run it with python bubble_sort.py.
Each model is tuned for one Indian problem and ships in GGUF or ONNX — small enough to run fully on-device, so the data never leaves the machine.
Detects and redacts Indian personal data — Aadhaar, PAN, phone numbers, Hinglish names — before text leaves a device.
Classifies scam and fraud SMS — KYC traps, fake offers, phishing — offline, on the phone itself.
Reads messy Indian bank statements and returns clean, structured JSON — private enough for real financial documents.
Farm advisory grounded in Kisan Call Centre data — crops, pests, subsidies — answering offline in four languages.
Scene-text recognition for Kannada — the OCR engine behind our document tooling for government forms.
Every model, dataset card and GGUF build, open on Hugging Face.
huggingface.co →They're why the models are small, the languages are local, and the billing is prepaid.
Aadhaar numbers, bank statements, medical papers — data like this should never need a cloud. If it can run locally, it ships locally.
Kannada, Telugu, Tamil, Hindi — trained on real Indian data like Kisan Call Centre logs, not translated afterthoughts.
India runs on UPI and prepaid balance. Usage is metered in rupees against a hard-capped wallet — no cards, no surprise bills.
A desktop app for Mac, Windows and Linux that gives you Claude and leading open models with your files kept on your own machine — funded by UPI top-ups, spent rupee by rupee.