Networking sold outSales open until 15.09 · 23:59--d --h --m --s
Web standards

AI-Transformation: Building AI-tools for Business Using the Newest Technology Breakouts

September 5, 2023 1:45 PM
Power Talk
🇬🇧 English
CMYK

About

We've been living in the digital transformation era for a really long time. From early 2023, we're living in the AI-transformation era. In this speech, we will discuss how to implement cutting-edge business AI-tools, their limitations, and the process that leads to great products.

Speakers

Photo of Marcin Fortuniak, Backend Team Leader at G-Group.dev
Marcin Fortuniak
Backend Team Leader at G-Group.dev
Photo of Michał Kurkowski, COO & Tech Lead at G-Group.dev
Michał Kurkowski
COO & Tech Lead at G-Group.dev

Watch the full talk

Watch this WaysConf session, then continue with related talks or explore the current programme.

Explore the 2026 agendaMore WaysConf recordings
From the recording

Talk in brief

At WaysConf 2023, Marcin Fortuniak and Michał Kurkowski from G-Group.dev explain how Foundational Model Operations (FM Ops) enables companies to rapidly build AI-powered business tools using existing large language models. Historically, custom AI development required massive datasets, specialized engineering talent, and high capital investment, resulting in brittle single-purpose models. The emergence of pre-trained models like GPT shifts the development focus toward orchestrating applications using pre-existing foundational infrastructure. The speakers outline the practical architecture of modern AI products, emphasizing model selection based on cost and task constraints, orchestration via LangChain, memory integration through vector databases, and connections to enterprise platforms like Jira, SAP, and Slack. They differentiate between predictable deterministic chains and reasoning AI agents that use specific tools to solve complex, unstructured business tasks such as customer support automation. Furthermore, they demonstrate how vector embeddings grant models access to internal company knowledge bases without requiring full model retraining.

Key takeaways

  1. 01

    Overcoming Historical AI Adoption Barriers

    Building custom AI models previously required extensive training data, rare machine learning talent, and massive financial investment. Modern foundational models eliminate these hurdles, allowing companies to integrate advanced capabilities without training models from scratch.

    Watch from 1:01
  2. 02

    Accelerating Development with FM Ops

    Foundational Model Operations (FM Ops) enables teams to build specialized software around existing large language models instead of training proprietary systems. This approach provides rapid prototyping, lower entry barriers for developers, and significantly reduced infrastructure costs.

    Watch from 5:54
  3. 03

    Matching Large Language Models to Task Requirements

    Organizations should select specific large language models based on task suitability and cost efficiency rather than defaulting to the largest model available. Surrounding the model with frameworks like LangChain, vector database memory, and external software integrations enables complex business workflows.

    Watch from 10:52
  4. 04

    Solving Unpredictable Tasks Using AI Agents

    While deterministic chains handle linear step-by-step processes, AI agents use internal reasoning to evaluate tools and resolve unstructured business scenarios. Equipping agents with contextual tools allows them to autonomously handle domain-specific workflows like customer support ticketing.

    Watch from 14:24
  5. 05

    Unlocking Proprietary Knowledge via Vector Databases

    Vector databases give foundational models long-term memory by converting enterprise text into semantic numerical embeddings. Searching these chunked embeddings enables AI tools to retrieve exact corporate knowledge without requiring model retraining.

    Watch from 16:42
Read edited transcript highlights

These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.

Overcoming Traditional AI Development Obstacles

Historically, small and midsize companies faced three major obstacles when attempting to build custom artificial intelligence models: exorbitant project costs, a severe shortage of specialized machine learning experts, and the requirement to assemble massive datasets for training from scratch. Even when organizations succeeded in funding and building a dedicated model, the resulting software was typically narrow in scope and could not easily be adapted to different business contexts or industries without incurring substantial additional development expenses.

Watch from 1:01

Understanding Foundational Model Operations

Foundational Model Operations, or FM Ops, represents a paradigm shift in software engineering centered around general-purpose models like GPT rather than hyper-specialized single-task models. Instead of spending months training networks on custom hardware, developers construct system architecture around pre-trained foundation models. This approach grants organizations blazing-fast prototyping capabilities, drastically lowers the technical barrier to entry for engineering teams, and maximizes the probability of successfully launching functional AI applications into production.

Watch from 5:54

Architectural Foundations of LLM Applications

Building robust enterprise AI tools requires selecting the right base model for a specific task based on performance and cost efficiency rather than defaulting to the largest available option. Developers then leverage orchestrating frameworks such as LangChain to connect the language model with external systems, databases, and enterprise platforms like Slack, Jira, or SAP. Integrating vector database memory enables the system to reference proprietary organizational data and historical context effectively during execution.

Watch from 10:52

Distinguishing Chains from Autonomous Agents

Within modern AI orchestration frameworks, workflows are divided into deterministic chains and autonomous agents. Chains execute fixed, predictable sequences of steps—such as querying a database and formatting the response—when a process is clear and linear. Conversely, agents are equipped with reasoning capabilities and specialized tool scripts, allowing them to dynamically evaluate complex scenarios, make decisions, and select the appropriate actions required to resolve open-ended business problems like customer service complaints.

Watch from 13:00

Leveraging Vector Databases for Semantic Context

Large language models naturally lack access to an organization's private corporate data and historical context. Vector databases solve this limitation by converting unstructured enterprise documents into numerical embeddings that capture semantic meaning. When a user submits a query, the system converts the input into a matching embedding, performs a vector similarity search across chunked internal records, and injects the retrieved context directly into the model's prompt to generate accurate, source-backed answers.

Watch from 16:42
Explore WaysConf 2026

View the agenda

Get tickets