Skip to main content
AppForge Solution - Webfejlesztés, Appfejlesztés, MI Fejlesztés
Back to home

Artificial Intelligence Portfolio

A showcase of our delivered AI projects. Every solution is backed by a real business challenge, custom development, and measurable results – from concept to production deployment.

01 RAG & Knowledge Base

Hungarian Pop Culture Archive – Intelligent Search Engine

The Hungarian Pop Culture Archive is one of the country's most comprehensive digital cultural repositories, documenting decades of Hungarian pop culture history across thousands of documents, books, and multimedia materials. The sheer scale and contextual complexity of the archive demanded an intelligent search system that goes far beyond traditional keyword matching.

The Challenge

Conventional search engines failed to handle the contextual richness of cultural materials. Users ask complex, natural-language questions – seeking connections between eras, genres, and creators – that cannot be answered with simple text matching. The organization's data governance policies required the entire system to run locally, without any external cloud service involvement.

Our Solution

We developed a locally deployed LLM and embedding system that vectorizes the entire document archive and enables semantic search. The RAG (Retrieval Augmented Generation) architecture ensures the model generates answers based exclusively on verified documents, minimizing hallucination risk. The system can synthesize answers from multiple documents and attaches source references to every response for traceability.

Key Results

  • Thousands of documents and books successfully indexed and vectorized
  • Natural language search with full Hungarian language support
  • Contextual answer generation with automatic source references
  • 100% local deployment – full data sovereignty guaranteed
  • One of the first production RAG systems demonstrated in Hungary

Technologies Used

Python FastAPI LlamaIndex LangChain Ollama ChromaDB React
02 Audio Analytics & AI

Real-Time Radio Monitoring & Analytics Platform

A real-time, 24/7 radio monitoring and analytics platform capable of simultaneously tracking broadcasts from multiple radio stations. The system uses advanced AI algorithms to automatically distinguish between content types – advertisements, music tracks, live programs, and podcasts – and generates detailed analytics, sentiment analysis, and statistical reports for each segment.

The Challenge

Radio market participants lacked real-time visibility into competitor broadcasts, running advertisements, and program sentiment trends. Manual monitoring was prohibitively expensive, slow, and subjective. A fully automated system was needed – one that could reliably monitor, categorize, and analyze radio content around the clock using objective metrics.

Our Solution

We integrated advanced speech recognition and audio classification models into a unified, scalable platform. We employ a fine-tuned Whisper model trained on over 500 hours of audio data, dramatically improving transcription accuracy. The combination of our fine-tuned Whisper and large language models (Gemini, Deepseek) enables automatic content segmentation, sentiment analysis, and thematic categorization. Results are displayed on an intuitive dashboard with real-time statistics, trend visualizations, and exportable reports.

Key Results

  • 24/7 real-time radio broadcast monitoring across multiple stations simultaneously
  • Automatic content segmentation: advertisements, music, programs, and podcasts
  • Detailed sentiment analysis and mood tracking for every segment
  • Comprehensive statistics, trend reports, and visualizations
  • Customizable alerts and automated report generation

Technologies Used

Python FastAPI Whisper Gemini Deepseek React PostgreSQL Redis
03 AI Pipeline & Automation

Automated Content Maintenance System

A fully automated content maintenance platform developed to keep the Hungarian Pop Culture Archive current. Through a custom AI pipeline, the system continuously monitors relevant news sources, analyzes new information, and generates update suggestions: fresh descriptions, updated data points, and previously undiscovered relationships between entities.

The Challenge

The Archive manages a massive and ever-growing cultural database whose content becomes outdated rapidly. Keeping up manually with new releases, events, and changes was impossible. The editorial team needed an intelligent tool that could automatically identify content requiring updates and provide contextualized suggestions for revision.

Our Solution

We built a multi-stage AI pipeline that automates the entire content maintenance workflow. The system scans relevant news sources, identifies connected cultural entities, and generates intelligent update suggestions. It doesn't just rewrite texts – it also proposes new relationships, metadata, and categorizations. Editors can approve, modify, or reject AI suggestions through a clean admin interface.

Key Results

  • Automated news source monitoring and relevant content identification
  • AI-powered text regeneration and description updates
  • Entity recognition and automatic knowledge graph expansion
  • Dramatic reduction in editorial time spent on maintenance tasks
  • Continuously fresh, up-to-date, and consistent cultural database

Technologies Used

Python FastAPI LangChain LangGraph Langfuse React PostgreSQL
04 Document Processing & Database

Mapei Product Database & Layer System

A comprehensive product database and intelligent layer system developed for Mapei, one of the world's leading construction materials manufacturers. The project aimed to build a complete, structured product registry from product catalogs previously available only in PDF format, complemented by a layer system that manages product compatibility and application sequences.

The Challenge

Mapei's product documentation – tens of thousands of pages of technical data sheets, application guides, and compatibility matrices – existed exclusively in PDF format. There was no centralized, searchable product database. Professionals had to manually sift through documents to find the right product specifications, application instructions, or compatibility information. The long-term goal was to build an AI-powered advisory system, which required structured data as a prerequisite.

Our Solution

We developed an automated document processing pipeline that extracted product data, specifications, application parameters, and compatibility matrices from tens of thousands of PDF pages. We designed an intelligent layer system that manages product application sequences and compatibility. The complete solution is served by a comprehensive admin interface and a robust REST API – preparing the infrastructure for the future AI-powered advisory feature.

Key Results

  • Tens of thousands of PDF pages successfully processed and structured
  • Centralized, searchable product database built from scratch
  • Intelligent layer system for product compatibility and application sequencing
  • Robust admin interface and fully documented REST API
  • Data infrastructure prepared for the planned AI advisory system

Technologies Used

Python FastAPI React PostgreSQL Redis Docker
05 Predictive Analytics & AI Advisory

Aviation Accounting AI Advisory System

An artificial intelligence solution integrated into the internal system of an international aviation accounting advisory firm. The system combines Text-to-SQL, ReAct agents, custom embeddings, and statistical models to support consultants in interpreting tax regulations, developing tax optimization strategies, and generating accurate predictions.

The Challenge

Aviation accounting is an exceptionally complex domain where tax rules vary significantly by country, aircraft type, and operating model. Consultants must review massive, heterogeneous datasets to develop optimal tax strategies. Manual analysis is extremely time-consuming, error-prone, and the constantly evolving regulatory landscape means prior experience can quickly become outdated.

Our Solution

We equipped the system with Text-to-SQL capabilities, allowing consultants to query the database in natural language without SQL expertise. ReAct (Reasoning + Acting) agents perform complex, multi-step analyses automatically: gathering relevant data, applying applicable rules, and producing structured recommendations. Custom embeddings and statistical models deliver the most accurate predictions for tax liability calculations, while the system proactively suggests optimization strategies based on historical data and current regulations.

Key Results

  • Natural language database querying via Text-to-SQL technology
  • Autonomous AI agents for complex, multi-step tax analyses (ReAct)
  • Predictive models for accurate tax liability forecasting
  • Proactive, data-driven tax optimization recommendations
  • Significant reduction in consultant analysis time

Technologies Used

Python FastAPI LangChain LangGraph LangSmith Langfuse React PostgreSQL
06 Speech Recognition & Document Analysis

Archival Audio Materials AI Processing

Comprehensive digital processing of historically valuable archival audio materials using artificial intelligence. Local processing of old, often extremely low-quality MP3 recordings, including full text transcription, content analysis, timeline structure generation, speaker identification, word cloud generation, and sentiment analysis.

The Challenge

Archival audio materials are typically of extremely poor quality: background noise, low-quality microphones, varying volume levels, and outdated recording formats pose significant challenges. Multiple speakers appear in the recordings, and the context often requires specialized historical and domain expertise. Due to the sensitivity of the materials, all processing had to be performed entirely locally, and the output had to be professional, citable, and archivable analysis documents.

Our Solution

We developed a locally deployed speech recognition system built on a fine-tuned Whisper model trained on over 500 hours of audio data, specifically optimized for low-quality archival audio. Following transcription, large language models analyze the text across multiple dimensions: identifying speakers, building timeline structures, performing sentiment analysis, generating word clouds, and producing comprehensive summary reports. The result: a complete, professional analysis document is generated for every audio recording.

Key Results

  • Successful transcription of low-quality, aged recordings with high accuracy
  • Automatic speaker identification and segmentation
  • Timeline structure construction for every recording
  • Sentiment analysis, keyword extraction, and word cloud generation
  • 100% local processing with full data security guaranteed
  • Automatic generation of professional analysis documents

Technologies Used

Python FastAPI Whisper Ollama Kimi K2 React

Full Technology Stack

Python FastAPI Pydantic React LlamaIndex LangChain LangGraph LangSmith Langfuse Whisper Gemini Deepseek Python FastAPI Pydantic React LlamaIndex LangChain LangGraph LangSmith Langfuse Whisper Gemini Deepseek Python FastAPI Pydantic React LlamaIndex LangChain LangGraph LangSmith Langfuse Whisper Gemini Deepseek
Kimi K2 ChromaDB Qdrant Pinecone Ollama vLLM PostgreSQL Redis Temporal Celery Docker Kubernetes Kimi K2 ChromaDB Qdrant Pinecone Ollama vLLM PostgreSQL Redis Temporal Celery Docker Kubernetes Kimi K2 ChromaDB Qdrant Pinecone Ollama vLLM PostgreSQL Redis Temporal Celery Docker Kubernetes

PARTNER WITH US TO CREATE A COMPELLING NARRATIVE FOR YOUR BRAND!

Let's Talk!
Contact

Get in Touch

Tell us about your project and we'll respond within 24 hours!

Additional details (optional)

Thank you! We'll get back to you soon.

We usually respond within 24 hours.

CEO

Boncz Balint

Office

Budapest, Hungary

Let's Talk Live!

Book a free 30-minute consultation where we discuss the details of your project.

Book a Call