Custom web applications, automated processes and intelligent systems. Discover software that helps our clients handle complex tasks every day.
Client Projects
Realised projects
Live
trainCo – The platform for hands-on professional training
From accredited training programmes to executive coaching: trainCo has stood for hands-on professional development with personality for two decades. We translated the institute's expertise into a modern web presence.
A custom web app developed for the Viennese housing developer. It automates location-based administrative tasks in the background, runs with absolute reliability and is in use every day.
Modular timber construction demands highly precise, complex processes. We digitised the recurring ones, from document search to automated, technically sound email drafts in day-to-day operations. Under ongoing support since 2021.
Your entire company knowledge, searchable like a conversation. jukeep turns scattered quotes, contracts, manuals and notes into a shared memory that answers any question in seconds.
Answers solely from your own documents, always with sources, never with invented content.
Knows its limits: if the information isn't there, jukeep says so openly.
EU-hosted and GDPR-compliant, on request entirely on your own server.
What makes the search special
An ordinary search only counts where a keyword appears most often, and often returns the wrong thing.
jukeep instead understands what a document is actually about. It combines classic full-text search with semantic meaning search (vector embeddings), merges both via Reciprocal Rank Fusion and uses an aboutness ranking to recognise which document genuinely answers your question, even when the searched word doesn't literally appear in it.
A tool that analyses business questions and examines them from different logical angles. Built to break up entrenched thinking within a team and to prepare complex decisions.
QR ordering system and cloud POS for hospitality. Menus can be photographed; yoorda reads the content via image recognition (OCR) and translates it automatically into several languages. Ordering, POS and the digital menu run together in one system.
Our consumer product (B2C) in the field of AI image generation. A standalone project we use to test our technology stack live, under heavy server load.
15 years of e-commerce experience at Media-Saturn.
Before webse.at existed, our technological foundation was laid at the Media-Saturn headquarters. There we helped shape one of Europe's largest e-commerce systems, ran international brand stores for global corporations such as Apple and Samsung, and delivered processes at scale. What did we take from it? A deep understanding that digital systems above all need structure, performance and measurable results. Good design, for us, is the foundation the actual result is built on.
Research & development
From the workshop
A look into our repositories, from local language models to our own open-source tools.
Fine-tuning small language models for a purpose
A training environment for small language models that run locally and are fine-tuned to a narrowly defined subject area. It covers the whole process — from preparing the training data through to the finished model.
A learning platform serving several education providers from the same installation, whose data stays separated down to database level (multi-tenancy). Alongside it, a learning app in which the learning logic is hard-coded: the language model phrases the content but does not decide the learning path.
Detecting changes in official sources
A system re-reads official sources at regular intervals and reports when something changes in a defined place. The detection was tested against historical cases in which such changes actually occurred.
Comparing public EU data over time
Weekly full extracts of a public EU register make it traceable which entries disappear again over time. Using the same raw data, we also recalculated an official cost estimate.
eigenlag: how fast a data pipeline can cycle at all
eigenlag calculates the shortest possible cycle time for recurring data pipelines. The limit follows from the structure of the schedule itself (max-plus eigenvalue) — more computing power cannot undercut it.
Vector search, built from the ground up
A teaching project rebuilds semantic search (vector search) step by step in readable code — from splitting a text through to the finished index. That makes it traceable how a query ends up as a list of hits.
A tool from control engineering computes entirely in the browser. There is no server that inputs are transmitted to; the data stays on the user's own machine.
What project are you planning?
Tell us what challenges your company is facing or which processes you'd like to automate. We listen and give you an honest, well-founded assessment of what's technically feasible and economically sensible.