· 001 — The thesis · Kerala, India · fully remote
Malayalam AI,built in Kerala.
Most AI is built far from the people who use it: trained on other languages, run in someone else's data centre, and judged by benchmarks that never test Malayalam.
Araskova starts from the opposite end: Malayalam AI should be measured, built and run in Kerala. We measure how leading models fail on real Malayalam tasks, build the licensed corpus and benchmarks to fix it, and deploy models that run on-premise.
Araskova Labs is founder-owned and run by a fully remote team in Kerala, established in 2026. We focus on Malayalam AI. Araskova Studio builds custom software for clients, and studio work funds the research. Araskova Academy is our education side: classes, hackathons, workshops and talks for students (@araskova_academy).
Malayalam computing has a long open-source history: volunteer communities such as Swathanthra Malayalam Computing (സ്വതന്ത്ര മലയാളം കമ്പ്യൂട്ടിങ്) built many of the fonts, input methods and language tools that Malayalam on computers relies on. We acknowledge that work; naming it here is not a claim of partnership.
- Malayalam corpus · Tokens
- 47.5M
- Malayalam corpus · Tokens
- Malayalam + English · as of June 2026
- Benchmark · Queries
- 200
- Benchmark · Queries
- Malayalam administrative tasks · 4 frontier models
- Deployment · By design
- On-prem
- Deployment · By design
- Runs on infrastructure the institution controls
- Established · Kerala
- 2026
- Established · Kerala
- Founder-owned · fully remote
· 002 — Origin
Six centuries ofKerala mathematics.
Kerala has a six-hundred-year record of original mathematics. Building Malayalam AI here is not cost arbitrage: the language, its speakers and its records are here.
· 003 — Architecture
Research. Systems.Products.
Research tells us where models fail on Malayalam; systems make the fix run on an institution's own hardware; products put it in front of the people who use it. One team does all three.
· 004 — Doctrine
How wework.
Five rules we hold our research, our systems and our client work to.
- 01
HYPOTHESIS FIRST
“Empirical baseline before line one.”
We write no production code without an explicit hypothesis, a measurable baseline, and defined failure boundaries. Speculative complexity is the enemy of reliability.
- 02
DATA SOVEREIGNTY
“Data stays with its owner.”
Records, recordings and model weights stay on infrastructure the institution controls. We design our systems to run on-premise, without sending data to outside services.
- 03
RUNS ON MODEST HARDWARE
“If it needs a data centre, it isn't finished.”
Offices, schools and archives run on ordinary servers and workstations. We size, quantize and test our models to run on that hardware, not on rented GPU clusters.
- 04
WORKING SOFTWARE
“Running binaries over slide decks.”
We do not sell mockups or roadmaps. Every engagement delivers working software, the benchmarks we measured it with, and the source code.
- 05
OBSESSIVE CRAFT
“Reliability is the point.”
We test against real inputs, handle failure explicitly and fix what breaks. Software should be predictable for the people who depend on it.
· 005 — What we ship
Studio andopen source.
The Lab's Malayalam document AI and assistants are on the products page. These are the systems Araskova Studio builds and supports, and the tools we release as open source.
· 006 — Fellowship
Founder-led · independent
Founder-led.Independent.The cohort of builders.
Araskova is founder-owned. We welcome partners and funders who back long-term Malayalam AI research.
Through the Araskova Research Fellowship, students work with our team on Malayalam NLP, evaluation and document AI. Fellowships are currently unpaid (15–20 h/week, 3–6 months).
· 007 — Partners · institutions · clients