---
title: "How to Teach Your Own AI Model, Part 2: What Should You Build First?"
description: "Explore practical AI model training, from choosing a first project and preparing data to fine-tuning locally and evaluating synthetic data."
url: https://www.callstack.com/events/how-to-teach-your-own-ai-model-part-2-what-should-you-build-first
updated: 2026-10-24T08:39:20.130Z
---

# How to Teach Your Own AI Model, Part 2: What Should You Build First?

Explore practical AI model training, from choosing a first project and preparing data to fine-tuning locally and evaluating synthetic data.

_Livestream · 7/16/2026 · Online · Recording: 02:27:01 · Category: AI_

Organizer: Callstack

Speakers: Kewin Wereszczyński (Software Engineer @ Callstack), Piotr Miłkowski (Senior AI System Engineer @ Callstack)

[Watch the recording](https://www.youtube.com/watch?v=7GT22LURIaY)

This session brought Kewin and Piotr back together to explore the practical side of teaching your own AI model: choosing a project, preparing the data, training a model, and understanding the trade-offs behind the tools.

## Start with a model you can understand and run

The conversation opened with the rapid progress in [smaller, more efficient models](#yt-time=00:08:11). Piotr unpacked model quantization through recent examples, showing how ternary and binary representations can dramatically reduce a model's memory footprint while retaining much of its quality. The discussion focused on what this shift means in practice: increasingly capable models can run locally, including on consumer hardware and mobile devices.

## Choose a focused problem and prepare the data

Training starts with a well-defined task and data that matches it. Kewin and Piotr [discussed](#yt-time=01:50:15) practical starter projects across computer vision, natural language processing, audio, video, and structured data. They also demonstrated how annotation tools such as Label Studio make the data-preparation process more tangible: labeling examples reveal what a model needs to learn and where a dataset may fall short.

## Fine-tune without losing sight of the fundamentals

The stream covered accessible [tools](#yt-time=01:59:36) for experimenting with fine-tuning, including Unsloth Studio, while connecting them to the underlying libraries many developers will encounter, such as PyTorch and Transformers. A graphical interface can lower the barrier to experimentation, but understanding the model, the training data, and the parameters remains essential to evaluating the output.

## Treat synthetic data with care

[Synthetic data](#yt-time=02:05:51) can extend a dataset, particularly when real examples are limited. The speakers also examined its limitations: errors in generated data can carry into later models, reducing quality over time. Their advice was clear: use well-described, high-quality datasets, and keep human expertise close to the data and evaluation process.

## Learn by building

For developers who want to go deeper, Kewin and Piotr [recommended](#yt-time=02:13:39) starting with AI fundamentals, learning the basic mathematical concepts behind models, and then building small projects. Running a model locally, fine-tuning a focused use case, or creating a LoRA for image generation can turn abstract concepts into practical understanding.