model status: deployed in production

A language model
trained on real life.

Prompt Engineering Linguistics PhD Synthetic Data Applied ML EdTech AI Creative AI
↓ explore training pipeline
Training Pipeline
Four phases. Each building richer representations of language — and what to do with it.
01 pre-training
🇹🇷

Turkey → Linguistics

Training data: Turkish · English · language fundamentals

Initial weights loaded in Turkey. Early epochs spent absorbing the deep structure of an agglutinative language — learning that word order isn't just syntax, it's information architecture.

100%
02 fine-tuning
🎓

PhD · UCLA Linguistics

Epochs: 2014–2020 · Loss fn: dissertation defense

Specialized training on Turkish syntax and Information Structure. Dissertation explored how speakers encode meaning through word position. Supervised by Profs. Mahajan & Sportiche. Taught LING 20.

100%
03 alignment
📚

AI Labs · ETS

Princeton, NJ · ~3 years · #AIforgood

Calibrated for educational impact. Part of AI Labs since day one — building data pipelines, synthetic data, and ML for assessment at scale. The model learned alignment: making capabilities serve human goals.

100%
04 deployment
🚀

Applied AI · Spotter

Since 2024 · Inference mode: creative generation

Deployed in production. Integrating generative AI into Spotter Studio for YouTubers — video ideas, thumbnails, content planning. Collaborating across Front-End, R&D, and Analytics. Live at scale.

82%
Model Architecture
Technical specifications of the SÖG-1 model.
Base Model
Linguistics
Syntax, semantics, pragmatics, information structure
Attention Heads
3 Languages
Turkish · English · the grammar of problem-solving
Context Window
∞
From agglutinative morphology to YouTube thumbnails
core_layers
Prompt Engineering Synthetic Data Applied ML Data Engineering Computational Linguistics
tool_layers
Python LLM APIs Data Pipelines Generative AI EdTech Creative Tooling
Benchmark Results
Performance metrics across key evaluation dimensions.
PhD
UCLA Linguistics
~3yr
AI Labs · ETS
0
M+ Creators Served
0
Industries Changed
Research Papers
Published outputs from the training process.
Dissertation · 2020

Syntax and Information Structure in Turkish Word Order Variations

PhD Dissertation · UCLA Linguistics

A comprehensive investigation into how Turkish speakers encode information through syntactic position — revealing the deep connection between word order flexibility and pragmatic meaning.

Publication · 2013

The Syntax of Multiple Internal Arguments in Turkish

Academic Publication

Early research exploring argument structure in Turkish — foundational understanding of how agglutinative languages handle complex predicate-argument relationships.

Conference · IEEE

Building Scalable AI Solutions

IEEE Conference Speaker

From theoretical linguistics to production AI — sharing lessons on building scalable solutions that bridge research and real-world deployment.

Teaching · UCLA

LING 20 — Introduction to Linguistics

UCLA · Instructor

Knowledge distillation from a specialized model to hundreds of student nodes. Teaching the next generation how language works — from phonetics to syntax.

"The best language models understand context, nuance, and when to generate something unexpected. The same is true for careers."
— the SÖG-1 training log, probably
🎗️
Cycle for Survival
West LA rider for rare cancer research at Memorial Sloan Kettering. Even deployed models give back to the training community.
🎤
IEEE Speaker
Open-sourcing production insights back to the research community. Building bridges between academia and industry.
🌍
Cross-Domain Transfer
Turkey → UCLA → Princeton → LA. Each migration expanded world knowledge and ability to generalize across domains.