aboutteamprojectsblogschat
aboutteamprojectsblogschat
© 2026 ottomind.ai

blogs.

How Can a Language Model Be Made Better in a Specific Domain?

A general-purpose code model does not preserve the same quality in a narrow and specialized domain. So we tried to make Code Llama perform better on the Unity side by extracting data from Unity documentation, turning it into question-code pairs, and fine-tuning the model. The result was clear: fine-tuning really made a difference. What created the real shift was not only choosing a good model, but feeding it the right domain data and pushing it closer to the right context.

18 min read · Mar 12, 2026

How Unstructured Data Stays Truly Accessible?

In archives, the core problem is often not missing data, but data that was not stored in a way that makes it usable later. Keeping the file does not mean accessing the information inside it. In this piece, we look at how unstructured data should be prepared, where it should be split, how its structure should be preserved, and which metadata should support it. What makes data valuable is not storing it, but making it truly accessible later.

16 min read · Dec 04, 2025

COREOCR: End-to-End Document Processing for Arabic Handwriting

COREOCR is an end-to-end document intelligence system we built not just to read Arabic handwriting, but to turn it into structured and usable digital data. It addresses challenges such as connected character structure, diacritics, irregular line flow, and data scarcity through transformer-based recognition, line processing, structural output generation, and an LLM-supported final correction layer. With diverse data sources, synthetic data, and a multi-stage training approach, it establishes a strong core that can operate in real document workflows.

15 min read · Sep 24, 2025

How to Choose the Right Examples in Few-Shot?

Using few-shot alone is not enough. In this piece, we looked at the difference between giving examples randomly and choosing them deliberately. What we saw was clear: even selecting better examples with a fixed logic changes the result. But the strongest effect comes from the dynamic approach, where we reselect the most semantically relevant examples for each question. In other words, the real difference comes not from few-shot itself, but from the quality of the choices inside it.

10 min read · Jul 17, 2025

What Changes Between Zero-Shot and Few-Shot?

The same language model can produce different results under different contexts, even on the same task. In this piece, we looked at the difference between zero-shot and few-shot, and at how the examples we give the model can change the answer. What we saw was clear: the issue is not only what the model knows, but how we explain the task to it. Context is a much more decisive layer than it first appears.

13 min read · May 28, 2024