NEST

Company

From a university project to a company

NEST started as a reading list and an architecture sketch at Istanbul Commerce University in February 2026. Seven months later it is a working codebase, one withdrawn number, and a decision to incorporate. Here is the mission, the timeline, the people and how to reach us.

Mission

Two things worth building

Give people who cannot speak or type a non-invasive path to open text: no surgery, no fixed phrase set, a decoding layer that runs on any research-grade EEG system. And give the field a benchmark it can trust, in an area where published scores have repeatedly turned out to reward the wrong thing once someone checked them properly, including a number NEST itself had to withdraw.

Story

Seven months, in order

Feb 2026
Project startsNEST begins at Istanbul Commerce University: literature review of EEG-to-text work and the first architecture design.
Feb 2026
First training runThe ZuCo corpus is integrated. An ad hoc LSTM decoder trains for 100 epochs on 12,071 real samples; loss decreases. A word error rate estimated from that loss, not from decoding, is published, then withdrawn.
Feb-Apr 2026
v2 and honest evaluationA new transformer architecture, NEST v2; an evaluation path built around free decoding instead of teacher forcing; this website.
Sept 2026
Decision to incorporateDecision to form a company. Applying to Hasat 2026 (Teknopark Istanbul), deadline 21 September 2026.
Q4 2026
Planned: validated benchmarkGPU training of v2 through to completion, with a free-decoding word error rate reported next to a noise-input control.
2027
Planned: corpus, incorporation, pilotsAn ethics-approved Turkish reading corpus, formal incorporation, and the first research pilots.

Principles

What we will not compromise on

Measure before claiming

A February 2026 estimate had to be withdrawn once the team compared it against real decoding, so nothing ships now until it has been decoded, not inferred from a loss curve.

Open by default

Code, data loaders and evaluation scripts are MIT-licensed and public, so any claim on this site can be checked instead of taken on faith.

Hardware-agnostic

NEST is a software layer, not a headset, so any research-grade EEG system that can output word-aligned frequency-band features is a valid input.

Research tool first, clinical product later

Making no diagnostic or therapeutic claim keeps NEST outside FDA, EU MDR and Turkish TITCK device scope today, and a future assistive product would take the CE/MDR path deliberately, not by accident.

Team

Who is behind it

One founder today. The first hires come after the first benchmark, on the clinical and regulatory side.

Hasan Tatar
Founder

Machine learning and software. Computer Engineering, Istanbul Commerce University. First author of an IEEE SIU 2026 paper on cross-lingual evaluation of vision-language-action models (IEEE Xplore 11636432). Founder of CamelLayer.

Istanbul Commerce University
Academic setting

NEST started as a research project at the university. Collaboration on ethics approval, participant access and EEG recording is planned; the team grows after the first validated result.

Ecosystem

Who NEST works with

Istanbul Commerce University

Home institution. Planned ethics board review, participant recruitment and EEG recording collaboration for the Turkish reading corpus.

Teknopark Istanbul Cube incubator

Planned incorporation venue.

TUBITAK BiGG

Planned grant application.

Hasat 2026

Application in preparation, September 2026, Biotechnology and Health track, hosted by Teknopark Istanbul.

Contact

Reach us

No mailing list and no support email yet. Use GitHub for anything code- or research-related, and LinkedIn for everything else.

Watch the company take shape

Star the repository for the Q4 2026 benchmark, or reach out on LinkedIn.