In February, Bitmovin launched the AI & Beyond Internship Incubator program designed to promote young talent in AI. The program is open to students from HTLs and technical universities to gain practical experience with AI-powered software solutions.
For the past few weeks, the six students joining the inaugural AI & Beyond Internship Incubator on July 1st have been working with Bitmovin’s technology experts on meaningful real-world AI projects within the product development teams, receiving mentorship along the way.
In this blog, we take a look at the impressive progress the students have made so far as they presented their work at the midterm at the end of July.
Midterm highlights from the program
Launched in March, Bitmovin’s AI Scene Analysis (AISA) is a feature that uses AI to analyze video content at the scene level, generating metadata about the content’s visual, audio, and narrative elements.
During their internship, Amelie Guger and Kristiyan Boyanov have been working on a joint project aimed at taking AISA to the next level, enabling customers to search, categorize, and repurpose content at the shot level in multiple languages.
For her part of the project, Amelie has focused on multiple language outputs. Currently, AISA’s primary language output is a JSON in English. Using LLMs to analyze and summarize individual video shots, AISA can now automatically translate those summaries into languages such as German and Japanese. Amelie implemented the translation pipeline using Gemini 2.0 Flash to balance quality and cost.
Kristiyan has been looking at LLM’s ambiguity within AISA and building a Schema to turn an unpredictable model into a predictable, dependable collaborator. This course of action was identified after a scene analysis summary in the JSON file included a paragraph about immunotherapy, despite the scene not featuring this anywhere at all. Now the algorithm should be able to predict any ambiguity, so it doesn’t appear in the scene analysis summaries.
Thanks to their joint work, the system now supports:
- Searchable, timestamped metadata per shot
- Multilingual summaries are shown directly in the Bitmovin VOD Encoder Dashboard using a dropdown to choose the language
- A foundation for automated vertical clip generation based on natural language prompts
Real-time anomaly detection in streaming analytics
Doru Floare has been working in the analytics department on developing a real-time AI model that can be used for anomaly detection in data. Based on LSTM (Long Short-Term Memory), which was chosen for its human-like qualities, the model has been trained on time series data, including timestamps and number of video plays to create a system that not only flags irregularities but adapts dynamically by understanding the difference between a true issue and a natural spike or recovery.
Key innovations include:
- Adaptive thresholding to reduce false positives after big spikes
- Log scale transformation for better sensitivity across different data volumes
- Root-cause analysis to pinpoint platform-level anomalies
Next for Doru is expanding the model to detect anomalies in metrics such as buffering and error rates with a focus on live alerting and deeper insight generation. He is hopeful that the Bitmovin team will be getting the first live alerts from the system very soon.
Error log tool with LLMs for support engineering
When an encoding fails, it often requires teams to sift through error logs from multiple sources, which takes time and slows down time to resolution (TTR). Kerem Gürtler has developed an LLM-powered tool that filters, deduplicates, and summarizes encoding logs, providing support engineers with clean, readable analysis to quickly pinpoint root causes. The demo highlighted:
- Integration with Bitmovin’s Support Copilot – ALPHA interface for natural language queries
- Encoding phase summaries
- Identification of recurrent or subtle errors that humans may overlook
The tool is already helping teams debug faster. The next steps include benchmarking accuracy, adding support doc integration, and extending the system for follow-up questions. Kerem is also working on a second project featuring genetic test cases using LLMs for test automation.
AI-powered industry insights reporting
Lukas Hirsch has addressed the challenge of exploring and interpreting platform-wide streaming trends by creating a natural language interface for querying a newly built Industry Insights Dashboard for Bitmovin’s Analytics. The feature could potentially allow customers to take a deep dive into the performance of products across platforms, look at stream time across live, video, or on-demand, and even geographical regions. In his demo, Lukas shared the intuitive dashboard he has created that displays a series of charts created from the questions asked. For example, a query such as ‘I want to see the error rate per session for different regions’ can be entered. To make this customer-facing, Lukas did acknowledge data privacy as a consideration for a wider rollout.
Plans include a chat-based interface to support open-ended questions and the potential for agentic AI to deliver results without requiring a specific query up front. For example, rather than asking ‘what is the startup time across different devices,’ a query could be ‘tell me what’s interesting about my startup times’.
Anton AI – Executive Assistant Agent
In the final presentation of the day, Michael Bumbar provided the gathered audience with an entertaining but compelling idea by creating an AI-powered assistant to replace the CEO, a bold idea given that Stefan Lederer was in the room. However, as Michael went through his demo, it was clear how the multi-agent system had been designed to act as a personal assistant rather than a replacement for Bitmovin’s leadership team.
Welcome, Anton AI, which uses agentic AI to carry out daily tasks such as drafting emails and summarizing Slack conversations. Acting as the ‘orchestrator agent,’ Anton can communicate with the ‘planning agent’ to carry out multi-step plans on command. Supported by the ‘workflow agents’ who can search Slack, Gmail, Search, and Calendar, tasks can be performed from a series of prompts. The demo highlighted:
- A natural language planning interface with optional step-by-step approvals.
- Access to internal tools like Google, Slack, and Gmail.
- A modular design allowing future expansion with open-source models.
In the second half of the internship, Michael plans to enable Anton to join meetings, transcribe calls, and learn over time from conversations. However, one audience member joked that Anton would have to join meetings late, because if he arrived on time, they would know it wasn’t Stefan!
At the midterm, it shows that innovation thrives when bright minds are given the tools to explore. Each project is already delivering value and pointing toward exciting futures.
“Each of these projects is more than an academic exercise; they are real, practical solutions already influencing Bitmovin workflows and products. Whether it’s smarter monitoring, faster support, deeper industry intelligence, or better content tools, our interns are helping define what AI at Bitmovin can (and should) look like,” Stefan said as he wrapped up the midterm session.
Final reflections and next steps
Next for the students is their final presentation on September 1st in Vienna, where they will all showcase their final projects. The AI & Beyond Internship Incubator Program for 2026 is now open. Register interests for the next intake in 2026: AI & Beyond 2026, or follow us on LinkedIn for more updates.
