Decision AI for Turkish, and AI that writes original exam questions.
Erenext LLC builds software products in its AI Lab: Pusula, an open-source Turkish decision model, and QuestNext Cloud Generator, now in development.
Pusula
Open-source Turkish decision model, Apache-2.0.
QuestNext
Autonomous exam-question generation. Not yet sold.
Bootstrapped
Founded 13 July 2026. Pre-seed. Team of two. $0 outside funding.
Turkish institutions and exam publishers are underserved by today’s AI.
Decision models are built for English
Widely used decision models were developed primarily for English. Their accuracy on Turkish is limited, especially on informal writing, typos, and mixed Turkish–English text.
Regulated data cannot hit public APIs
Banks, health providers, law firms, and public services often cannot send customer messages to a public AI API under Turkey’s KVKK personal-data rules. They need models they can run themselves.
Original questions are slow to write
Schools and publishers need original, curriculum-aligned exam questions. Writing a new item for each learning objective still depends on scarce educator time.
An open-source Turkish decision model that returns probabilities, not generated text.
Pusula answers typed questions about a text and returns a probability for every option. Typical uses: routing, urgency, human handoff, fraud, and prompt-injection checks.
- Built-in Turkish taxonomy: 20 domains, 103 subdomains, 333 intents, plus 20 ready-made templates.
- Pusula (GPU): 26B-class Decision 3.0 — about 64 GB in BF16, or 24 GB in 4-bit.
- Pusula Nano (CPU or any GPU): 149M parameters.
Turkish Decision Benchmark
Pusula 2.0 score, as published on the product page. We train and serve the model for institutions that cannot send customer messages to a public API.
Product: erenext.com/ai/pusula · Weights and code on Hugging Face and GitHub (erenextlabs/pusula)
Autonomous exam-question generation, now in development.
QuestNext turns topics and learning objectives into original multiple-choice and open-ended exam questions for curricula and admissions tests.
Covered curricula and tests include IB, AP, A Level, YKS, IMAT, SAT, ACT, ESAT, TOEFL, and IELTS.
Concept demonstration only
https://questnext.cloud/ is an interactive concept demonstration. It does not generate questions. The product is not yet sold.
Product page: erenext.com/ai/questnext
Shipped open-source software, a product in development, and a revenue-generating company.
Pusula is public
Open-source release on 31 October 2026, Apache-2.0, with a published 92.0 score on the Turkish Decision Benchmark.
QuestNext is not yet sold
The generator is in development. The public site is a concept demonstration only.
Profitable and bootstrapped
Founded 13 July 2026. Pre-seed, $0 outside funding, paying customers, revenue-generating and profitable. No outside capital.
We also apply the lab’s models to educator-checked, AI-generated video lessons. That line is a commercial application of the products, not a school.
Product-led, with institutions first.
Deploy, fine-tune, support
We train and serve Pusula for institutions that need on-their-own-infrastructure inference under KVKK: deployment, fine-tuning, and support.
Not yet commercial
QuestNext Cloud Generator is in development and is not yet sold. When it ships, it is intended as a product for curriculum and admissions-question generation.
Educator-checked lessons
Interactive video lessons for IB, AP, and A Level Physics, Chemistry, and Mathematics, produced with AI and checked by Erenext educators. A secondary custom-AI software line exists; Erenext is not a consultancy or a school.
Two people, building the lab and the products.
Hasan Şahin
Founder of Erenext LLC and Pusula project lead. The company is a US LLC registered in Denver, Colorado; Hasan is based in Turkey.
Ahmet Sayman
Second teammate, Developer/Engineer. The company is a team of two.
Isolated sandboxes for evaluation and training — planned, not live today.
Evaluation and training jobs
We would use Daytona for isolated evaluation and training jobs for Pusula, including checks that should not run on a shared public API.
Disposable generation pipelines
As a planned use, we would run and evaluate QuestNext’s autonomous question-generation pipeline in a disposable environment, without exposing institutional customer data or student personal data.
We are not already running this workload on Daytona. Intended use cases: AI evaluations and training & fine-tuning.