Ecosystem of AI Agents

Q2 2025.

This is an intelligent assistant for the admissions committee of the Kazakh-American Free University. It uses a local LLM and RAG system to find relevant answers to user queries.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 PostgreSQL database - Used to store embeddings (vectors) of responses split into text chunks;
• 2 Local model Qwen3-Embedding-0.6B - Multilingual model used to generate vector representations (embeddings) of texts;
  • Model context: 32,768 tokens.
• 3 Local model Google/Gemma-4-31B - Multilingual language model with tool_choice support, allowing connection of external tools such as Internet search and the ability to process data from the RAG system;
  • Model context: 7,500 tokens.
• 4 RAG and RAG-Fusion mechanism - Used to combine and rank results from multiple queries and form the most relevant context before generating a response by the model;
• 5 Reranking mechanism using the bge-reranker-v2-m3 model - Used to select the most accurate and relevant answer from multiple found fragments;
• 6 Voice recognition + RVC models - Operates based on the local Vosk model, providing speech-to-text conversion (currently in Russian) and an RVC server for generating human-like voice responses;
• 7 Final result - Passing the final message in the format:
  prompt copy

  {Context}: %Result%.
  {Question}: %Question%.
  {Note}: Give an answer to the applicant.

The agent is available 24/7 on Telegram and helps applicants, parents, and schoolchildren learn everything about admission to our university.

Q2 2025.

This is a library assistant that helps users find books, recommend new items, create tables and compare them, and automatically format bibliographic references according to GOST.

A solution from OpenAI in the form of API access to models is temporarily used. A transition to a local LLM will be made soon. Testing of filling tables on a local LLM is underway.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 MySQL database - After the integration of the automated library system KABIS, this agent gained the ability to query the book database to provide answers to user questions;
• 2 Using the RMEL and IPRSmart API - Querying resources to search and retrieve books;
• 3 Model gpt-4o-mini - Multilingual model that uses tool_choice to execute queries to the book database and output results;
  • Model context: 65,536 tokens.
• 4 Final result - The model answers basic questions related to the library. In cases where the user requests any books, the model chooses when to call functions, after which it outputs the final result to the user.

The library assistant is available on the university website and helps students, teachers, and researchers 24/7.

Q4 2025.

This is an AI platform for automated educational course creation, lecture recording with AI avatars and test generation.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 Video recording with teacher AI avatars - Built-in system for creating video lessons with realistic AI avatars. Avatars are generated based on photos and audio recordings of real teachers, creating a "live presence" effect and increasing student trust;
• 2 Generation of educational materials - The agent automatically generates lecture texts, presentations, teaching aids and practical assignments for a given discipline. It uses university standard templates and discipline databases;
• 3 Realistic AI avatars - Face generation and speech synchronization models with a high degree of realism (up to 95%) are used. This allows students to perceive content as a live lecture, which increases engagement and trust in the teacher;
• 4 Final result - The agent creates a ready-made multimedia course: includes video lectures, teaching materials, tests and links. The teacher only has to approve and publish the course;
• 5 Test creation - To generate tests in GIFT format and lecture materials, the local model Google/Gemma-4-31B is used
  • Model context: 9,500 tokens.
• 6 Image creation - To generate images, the local model Fooocus and the RTX 5090 32 VRAM graphics card are used.

The AI platform is available 24/7, allowing users to create high-quality videos with virtual avatars, speech synthesis and support for multiple languages, as well as lecture materials and tests at any time.

Q4 2025.

Digital assistant to help international students adapt to the Kazakh-American Free University. Answers questions about accommodation, visa support, academic process and student life.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 FastAPI - Web framework for building API and processing user requests;
• 2 Redis - Used for temporary storage of chat history;
• 3 Local model Qwen3-Embedding-0.6B - Creates embeddings and helps the system understand the meaning of queries;
  • Model context: 32,768 tokens.
• 4 Local model Google/Gemma-4-31B - Main language model for generating responses to users;
  • Model context: 7,500 tokens.
• 5 LangChain - Library that ensures interaction between local models and dialogue logic;
• 6 Final result - Personalized response to international students' questions:
  prompt copy

  {Question}: What educational programs are available?
  {Answer}: The following educational programs are available: …
  

The agent is available 24/7 and is hosted on the KAFU AI-Sana platform. Its goal is to facilitate the adaptation process for international students and make communication with the university more convenient and understandable.

Q2 2025.

This is an intelligent assistant for students of the Kazakh-American Free University. It uses a local LLM and RAG system to find relevant answers to user queries.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 PostgreSQL database - Used to store embeddings (vectors) of responses split into text chunks;
• 2 Local model Qwen3-Embedding-0.6B - Multilingual model used to generate vector representations (embeddings) of texts;
  • Model context: 32,768 tokens.
• 3 Local model Google/Gemma-4-31B - Multilingual language model with tool_choice support, allowing connection of external tools such as Internet search and the ability to process data from the RAG system;
  • Model context: 7,500 tokens.
• 4 RAG and RAG-Fusion mechanism - Used to combine and rank results from multiple queries and form the most relevant context before generating a response by the model;
• 5 Reranking mechanism using the bge-reranker-v2-m3 model - Used to select the most accurate and relevant answer from multiple found fragments;
• 6 Voice recognition + RVC models - Operates based on the local Vosk model, providing speech-to-text conversion (currently in Russian) and an RVC server for generating human-like voice responses;
• 7 Final result - Passing the final message in the format:
  prompt copy

  {Context}: %Result%.
  {Question}: %Question%.
  {Note}: Give an answer to the student.

The agent is available 24/7 on Telegram and helps students find out the necessary information about our university.

Q2 2026.

This is an intelligent assistant for students of the Kazakhstan-American Free University. A local LLM with tool_choice support is used for personalized selection of psychological tests.

Below are the main technologies and components that support the operation of the university's intelligent agent.

• 1 ASP.NET - Used as a microservice to manage the agent's operation;
• 2 MySQL Database - Used to store student requests for contacting a psychologist and additional information;
• 3 PostgreSQL Database - Used to store secondary information, including embeddings (vectors), such as university information or the psychologist's contact number;
• 4 NoSQL Database - Used for fast loading and temporary storage of chat history;
• 5 Local Whisper Model - A local multilingual ggml-large-v3 model is used for speech recognition;
• 6 Local Qwen3-Embedding-0.6B Model - A multilingual model used to generate vector representations (embeddings) of texts;
  • Model context: 32,768 tokens.
• 7 Local Google/Gemma-4-31B Model - The main language model used to generate responses to users, with tool_choice support for selecting psychological tests;
  • Model context: 9,500 tokens.
• 8 Final Result - The model answers basic questions related to psychological support and provides assistance. Depending on the user's request, it selects suitable psychological tests, gives recommendations, and, if necessary, offers to submit a request to contact a psychologist, after which it provides the final result to the user.

The agent is available 24/7 on the AI-Sana platform and is always ready to support students during difficult moments.

Q4 2026.

This is a public service for scholars and researchers in Kazakhstan. It answers two questions - whether it is safe to publish in a particular journal, and which journals are worth submitting an article to given its subject. Verification is carried out against official lists, while selection is based on the semantic proximity of the abstract to the journal's subject area, for which a local language model and vector search are used.

Below are the main technologies and components that ensure the operation of the agent.

• 1 FastAPI and SQLAlchemy - Used as a microservice with asynchronous database access through asyncpg;
• 2 PostgreSQL database - Stores the journal catalog, their membership in the lists, and their metrics; the pg_trgm extension enables search by an inexact spelling of the title;
• 3 pgvector extension - Stores embeddings in the halfvec type with an HNSW index by cosine proximity, to search tens of thousands of journals in milliseconds;
• 4 Data sources - DOAJ, Scopus (active and discontinued), OpenAlex, Scimago, Beall's List and the list of journals of the Quality Assurance Committee of the Ministry of Science and Higher Education of the Republic of Kazakhstan;
  • Updates: Automatic, on a schedule.
• 5 Local model Qwen3-Embedding-0.6B - Multilingual model used to generate vector representations (embeddings) of journal subject areas and of the user's abstract;
  • Vector dimension: 1,024.
• 6 Local model Google/Gemma-4-31B - Translates the abstract into English and expands a short query of a few words into a full abstract with methods and subfields - without this the vector becomes blurred and the selection loses accuracy;
  • Model context: 7,680 tokens.
• 7 KAFU Global API - A single access point to the university's infrastructure: /api/ai/v1/chat/completions and /api/ai/v1/embeddings. Researchers' abstracts never leave the university's perimeter;
• 8 React and Vite frontend - An interface built with TypeScript and Tailwind CSS, translated into Kazakh, Russian, and English;
• 9 Final result - A verdict on the journal with links to the sources, or a list of journals sorted by semantic proximity, with predatory ones excluded automatically and filters by quartile, open access, and the Quality Assurance Committee list.

The service is free and open to all researchers in Kazakhstan. It helps them avoid losing money and reputation on a publication in a disreputable journal and choose a journal that will count toward a dissertation defense.

Q4 2026.

This is a web platform for monitoring air quality in Kazakhstan for students, teachers, and residents of the East Kazakhstan Region. The whole country is shown on a single interactive map. A separate tab holds an AI agent that queries the data itself through tool_choice, searches the Internet, and builds charts right inside the conversation. The data is exposed through a public token-based API, so the index can be embedded in Moodle or on any university website.

Below are the main technologies and components that ensure the operation of the air monitoring platform.

• 1 Node.js and Express - Used as a microservice that collects the data, keeps a cache, and serves it to the interface and to external consumers;
• 2 Vite - Frontend build without a framework. The map, the region panel, the chat, and the API page are compiled into static files;
• 3 Interactive SVG map - Geometrically accurate borders of twenty regions, zooming, dragging, and highlighting of the selected region are drawn by hand, without any mapping libraries;
• 4 CAMS model through Open-Meteo - A calculated air quality estimate for a city point, with a resolution of about 45 km; it covers the entire country, including places with no stations;
• 5 Network of ground stations - Readings from city sensors on the US AQI scale; silent stations are discarded, otherwise they understate the city's average index threefold;
  • The response always shows what exactly came back: Stations or the model.
• 6 Local model Google/Gemma-4-31B - The AI agent's main language model, connected through the KAFU Global API with native support for tools and streaming generation;
  • Model context: 7,680 tokens.
• 7 Agent tools - The agent decides for itself which tool to use. It can check pollution and weather, compare cities, pull up a station's history, search the Internet, or build a chart;
  • Charts are drawn right in the conversation: Trends, bars, pollutant shares, and the weather forecast.
• 8 Streaming through SSE - The response arrives as it is generated. First the cards of the called tools unfold, then the formatted text; the send button turns into a square and stops the stream;
• 9 Public API with tokens - Any user can issue themselves a key of the form kafu-eco-tk-... and request data for any city in the country;
  • Only the SHA-256 of the token is stored on the server, and verification runs in constant time.
  • Issuance is recorded by IP, device, and cookie, and an administrator can revoke a token at any time.
• 10 Final result - The map, the conversation with the agent, and the public API all run on the same data, while a ready-made HTML block shows the current index right inside Moodle and refreshes it every hour.

The platform is open to everyone, the map can be viewed without registration, and a token for data access is issued in one click on the API page.

Q4 2026.

This is an agent for the departments and the educational and methodological office of the Kazakh-American Free University. It takes a working curriculum in .xls format as input and produces a completed form, "Information on the holdings of educational, methodological, and scientific literature", on Russian and Kazakh sheets. The literature is selected from the real holdings of the KAFU library by a local language model working through the tool_choice mechanism.

Below are the main technologies and components that ensure the operation of the agent.

• 1 FastAPI and Uvicorn - Used as a microservicefor receiving the curriculum, with a task queue and real-time progress reporting;
• 2 Curriculum parsing - The xlrd and openpyxl libraries extract the disciplines in two languages and the degree programme code from the title page;
• 3 SQLite database - A single local catalog of the holdings with a full-text index and vector representations of the books;
• 4 Literature sources - The KABIS (print holdings), IPR SMART and RMEB (electronic publications) catalogs, over 80 thousand titles in total;
  • Search: Full-text and semantic.
• 5 Local model Qwen3-Embedding-0.6B - Multilingual model used to generate vector representations (embeddings) of book titles and abstracts;
  • Vector dimension: 1,024.
• 6 Local model Google/Gemma-4-31B - The main language model, which formulates search queries, assesses how relevant publications are to the subject of the discipline, and returns the final selection through tool_choice;
  • Model context: 7,680 tokens.
• 7 KAFU Global API - A single access point to the university's infrastructure: /api/ai/v1/chat/completions for the model and /api/ai/v1/embeddings for the vectors. The data never leaves the university's perimeter;
• 8 Vite frontend - An interface with live progress of the selection by discipline and translation into three languages through JSON dictionaries;
• 9 Final result - A ready-made Appendix in .xlsx format with Russian and Kazakh sheets, plus an annotated version in which questionable publications are highlighted in red for review by a specialist.

Work that used to take a teacher up to a week for a single degree programme is done by the agent in about 5-10 minutes. All that is left is to check and sign the final form.

Q3 2026.

This is a voice AI assistant at the university entrance, running on a touchscreen kiosk for the students and guests of the Kazakh-American Free University. It runs on a local LLM with tool_choice support for tool selection and RAG, and the Whisper and CosyVoice3 models for speech recognition and generation.

Below are the main technologies and components that ensure the operation of the university's intelligent agent.

• 1 ASP.NET - Used as a microservice for managing the agent's operation;
• 2 Live2D engine - Used to bring to life the 2D render of a cat in university branding that serves as the university mascot;
• 3 OpenGL library - Used to render the cat and the entire interface on top of it;
• 4 PostgreSQL database - Used to store (embeddings (vectors)) and holds all the information needed to answer questions;
• 5 Local model Whisper - The local multilingual model ggml-large-v3 is used for speech recognition;
• 6 Local model CosyVoice3 - The local multilingual model CosyVoice3 is used to generate natural speech with emotions and pauses between punctuation marks, and it can use any voice through zero-shot voice cloning;
• 7 Local model Qwen3-Embedding-0.6B - Multilingual model used to generate vector representations (embeddings) of texts;
  • Model context: 32,768 tokens.
• 8 Local model Google/Gemma-4-31B - The main language model for generating responses to the user, with tool_choice support and subsequent voicing of the text;
  • Model context: 9,500 tokens.
• 9 Final result - The model uses streaming generation stream: true, feeding chunks of text into the input of the voicing model to answer questions from the university's students and guests.

The voice AI agent is located in the university and is always glad to help any student with their question.

Q3 2026.

This is an AI agent for auditing courses on the moodle.kafu.edu.kz portal, which compares the actual content of the courses against the approved teaching load of the departments. The agent is implemented as a specialized skill (skill) on the Claude platform.

Below are the main stages and components that ensure the operation of the course audit agent in the Claude environment.

• 1 The Claude skill KAFU AI Smart Audit - All the audit logic is packaged as a skill that is triggered by a request such as "check the department". The skill contains the rules for parsing the load, matching, and generating reports;
• 2 Parsing of teaching load files - The agent reads department files in the .xls, .xlsx formats and in archives, recognizing various layouts, and builds a list of "teacher - discipline - semester";
• 3 Crawling the Moodle portal - The agent retrieves the course catalog by department categories and teacher subcategories, reading the structure of the course pages directly from the browser DOM; when a web services token is available, the work runs through the Moodle REST API without any staff involvement;
• 4 Matching disciplines to courses - Fuzzy comparison of titles with text normalization and a confidence score; low-confidence matches are sent to a staff member for confirmation instead of being accepted automatically;
• 5 Comprehensive check of fifteen weeks - For each course the agent walks through all the study sections and records the presence of a lecture or presentation, an assignment or a quiz, determining the type of each element by its modtype_* class;
• 6 Reading the syllabus from the document content - The agent downloads the attached file and extracts the text from it, determining the year it was drawn up from the title page, the department's approval protocol, and the copyright line;
  • .docx - Unpacking the archive and reading document.xml and the headers and footers;
  • .doc - Extracting the text in the UTF-16LE and windows-1251 encodings;
  • .pdf - Layer-by-layer text extraction through pdf.js, joining up broken digits of the year.
• 7 Report generation - The agent assembles .xlsx workbooks in the approved format: a summary for the rector's office with the key indicators, a consolidated department table with links to the courses and a list of unfilled weeks, and individual sheets for each teacher;
• 8 Final result - In addition, a consolidated analytical report covering several departments is generated, with charts in the Word and PDF formats, ready to be presented to management without manual rework.

605 courses, 122 teachers, and 628 teaching load entries were checked across four departments: foreign languages, law and international relations, business, and pedagogy and psychology. The key conclusion of the pilot: determining the syllabus year from the file name overstates the result threefold - by file name 66-73 % of the syllabi looked current, whereas by document content only 18 % turned out to be.

The agent does not process students' personal data - only the structure of the courses and the educational and methodological documentation is analyzed. The functionality can be extended to the university's remaining departments as part of the development of the Smart Bilim initiative.