Discover our projects

The company recieved a large volume and variety of documents on a daily basis (invoices, delivery-acceptance notes, certificates of performance, etc.) Even within the same document type, there were significant variations, and due to the diverse range of partners, template-based solutions were nopt feasible. OCR solutions, however, were only effective when the information to be extracted consistently followed the same format.

Interprets documents and automatically extracts key information defined by the client.
Documents can be in various formats, including scanned files.

98% accuracy was achieved after training on the client's datasets.

Checking and reconciling orders, delivery-acceptance notes and invoice line items required significant administrative human effort before the introduction of generative AI. The process was typically slow and costly, especially when item names, units of measure, and other details varied across different documents.

Our solution, which can integrate with workflows and ERP systems, analyzes the content of documents and flags discrepancies. Users and/or managers can see these discrepancies before opening the documents.

The process became 30% faster and the number of human errors significantly reduced. This improvement was achieved by automating field population and providing discrepancy notifications, effectively supporting tasks that were previously manual, such as form completion and approvals.

Reports on workplace accidents are recorded textually in the corporate HSE system, however, they are often incomplete, do not meet company standards, and training the involved staff is very costly.
There are not enough specialist to process HSE reports, extract relevant information and summarize them, and their work time is expensive.

A solution was implemented to track and analyze energy data from building automation meters, providing actionable insights.

Report quality improves and the report authors gain a better understanding of their responsibilities through immediate feedback.
Processing and summarizing reports in the headquarters becomes significantly faster.

Analyzing resumes recieved in various formats is cumbersome for HR staff. Any new managerial request, filtering or search condition requires reopening the documents.
Traditional free-text search cannot interpret content or identify similarities, it only finds the exact text, often filtering out relevant results.

Iteratively analyzes and filters the document repository through questions asked in a chat interface.
Generates ideal candidate pofile based on the provided information (e.g., company culture).
Evaluates suitability of resumes according to the ideal profile.

A manual HR task that previously took 3-4 hours is completed in just a few minutes.

The foreign IT service company is successful at tenders, but there are too many simultaneously available tenders. Continuously monitoring, analyzing and selecting the most promising calls for proposals requires significant time and effort.

Automatically collects tender announcements available online
Applicant company profile can be configured by the user
Based on the applicant profile, the AI service analyzes and ranks tender announcements
Automatically notifies the user about the most relevant opportunities

The client automatically receives a personalized and prioritized list about the best tenders based on relevance

Employees often struggled to find answers in the company’s knowledge base, which resulted in managers spending time with routine question, diverting managerial time from higher-value activities

Integrated a chatbot within the ERP system used by the company
Vectorized company knowledge base
The chatbot responds based on internal knowledge or other sources if needed.
A modular system that can be easily extended with additional features, such as file processing or generation, as needed by the client.

… TBD, not ready yet

During the integration of a new line, the line needs to be tested, and the client requested that the process be automated and based on data analytics.

We trained a custom neural network on high quality samples and samples with audio issues.
The model is capable of evaluating an audio sample in terms of sound quality.
It does not store or interprets the provided audio file, making compliance with GDPR and AI Act requirements easier.
Upon request, we can train a model on the client’s own database, enabling differentiation based on various criteria (e.g., is the issue network-related or caused by the endpoint device?)

Fast, real-time or bulk analyis of audio files (e.g., archives).

There are not enough specialists to handle the volume of customer inquiries, making responses costly.
Responding is time-consuming and sometimes inadequate because searching for answers in the internal knowledge base is difficult.

Extracts relevant content from a large document repository.
Processes corporate documents of several hundred pages within seconds.
Generates a customized email response draft based on internal knowledge bases.
In addition to email and chat, the knowledge base can also be queried through voice-based communication.

Finding the required information becomes significantly faster, and response messages can be generated automatically.
The knowledge base can also be searched during live interactions, allowing operators to receive relevant information about the case in real time while communicating with the customer.

In customer service areas of banks and public administration offices, selecting the appropriate service can be difficult. Customers often require assistance from staff, which increases operational costs.

Based on the customer’s spoken request, the system automatically recommends relevant service options by selecting from an integrated service tree.
If the confidence level of the recommendation is low, the system asks clarifying questions to refine the suggestion.

The time required to select a service is reduced, and the need for assistance from human staff is significantly minimized.

At a large international company, customer churn has been steadily increasing in recent years. It is difficult to determine which customers should receive a retention offer and when it should be made in order to prevent contract termination. The goal is proactive intervention while using resources efficiently.

Construction of a time-series owner database with monthly granularity.
Development of a Cox time-varying churn risk model with dynamically changing features.
Continuous assignment of a monthly updated churn risk score to each owner.
Creation of risk groups (Low / Medium / High) and validation using retention curves.
Unsupervised clustering to identify distinct owner personas.
Conceptual design of a retention workflow and intervention loop, including preparation for uplift modeling.

Reduced customer churn through early, targeted interventions based on machine learning insights and data-driven retention strategies.
Improved overall customer satisfaction and more efficient allocation of resources.

A large corporation operates dozens of holiday parks across the UK. Planning staff levels at pubs, restaurants, leisure facilities and reception areas on a weekly basis is a major challenge and often far from optimal, especially since visitor numbers are strongly seasonal and volatile.
Decisions should be made at the venue level with hourly granularity, which is not feasible manually.
Due to the lack of revenue forecasts, procurement and inventory planning could not be properly optimized.

Footfall forecasting
Revenue forecasting
Staffing demand forecasting
Production pipeline and automation (MLOps on Snowflake)

Labor cost optimization, forecasting peak traffic periods, reduced understaffing and overstuffing
Instead of weekly experience-based estimation, 1-4 week forecasts at the venue and hourly levels

Placing order items onto pallets in the correct sequence requires careful planning, expertise, and precise knowledge of placement rules. Training staff for this task is costly, and despite this, errors still occur, leading to accidents, product damage, and suboptimal space utilization.

Based on the descriptive attributes of order items, customer-defined placement rules, and various optimization algorithms, the system determines the optimal loading sequence and the exact placement of each item on the pallet.
Can return multiple alternatives and optimize according to different criteria.

Warehouse staff receive exact task descriptions and visual guidance (2D-3D diagrams) for pallet placement. This accelerates the process and reduces errors.

A car-sharing company could only adjust vehicle prices manually and in a static way, based on retrospective data. The company needed an automated, dynamic pricing solution capable of reacting to both current internal and external factors.

Near real-time pricessing of large volumes of data.
Use of vehicle data combined with various environmental factors.
A flexible, rule-based, no-code system that can be easily defined and modified by the client’s staff
Support for complex rule definitions, discount system, ”what-if” scenario analyses

Within the first six months, service usage increased by more than 5% compared to the same period previously.
Automation significantly reduced the preparatory effort required to maintain and manage the discount system.

Support primary school students in their studies through an AI-based mobile application.

Significant R&D focused on AI-related components within the project
Mobil application with and AI-powered backend
Vectorized textbooks
Chat- and image-based interaction
End-to-end call chain architecture and LLMOps
Teljes hívási lánc kiépítése, LLMOps

The AI backend was successfully developed and integrated into the mobile application.