Rasa's 2025 Agent Building Challenge: Announcing Winner

We are excited to share the results of our nearly six-week-long Agent Building Challenge.

More than 450 participants applied to join the challenge from around the world: India, USA, Nigeria, Saudi Arabia, Philippines, Thailand, UK, Germany, Brazil, Slovakia, Turkey, Denmark, and many more.

We are thrilled by the number of ideas and use cases that participants came up with. They built AI agents across a wide range of domains, including dating, cryptocurrency, farming, elder care, support for women and young parents, navigation, rentals and housing search, and home renovation.

Judging Criteria and the Jury

All projects were judged based on the same criteria:

Is the project unique in its approach, and does it solve a specific problem in conversational AI?
How well the assistant can handle complex conversations (context switch, corrections, interruptions, etc).
How well does the assistant combine RAG and transactional flows?
Does the assistant have integrations with sophisticated backend systems?
Additional points: Does the assistant use open-source LLMs?
Additional points: Does the assistant use fine-tuned models?
Additional points: Does the assistant have integrations with other AI tools?

The jury was made up of people from varied backgrounds, each bringing their own unique experience:

Ben Epstein at MLOps Community
Ty Dunn, Co-founder and CEO at Continue
Justina Petraitytė, Developer Relations Engineer
Dr. Alan Nichol, Co-founder and CTO at Rasa
Daksh Varshneya, Senior Product Manager at Rasa
Marina Ashurkina, Developer Relations Engineer at Rasa

Announcing the Winners

🥇1st Prize - $10,000

Vladimir Preobrazhenskiy, Thailand – “AI Moto Expert Agent”, GitHub

Watch Vladimir's demo here:

Vladimir, a motorcycle enthusiast, built an AI agent that can recommend specific motorcycle models and answer a wide range of motorcycle-related questions; for example, which helmet to choose or which driving school to attend.

What stood out to us is how precisely the agent distinguishes between three very similar types of user questions, which is a common challenge in conversational applications:

General motorcycle questions (e.g., “What is ABS on a bike?”) are handled via semantic search. Vladimir generated question-answer pairs and vectorized them for similarity-based retrieval.
Model-specific queries (e.g., “Tell me about the Yamaha MT-07”) are answered using a local database containing structured information about around 7,500 different motorcycle models.
Open-ended or exploratory questions (e.g., “Which driving school should I choose in Bangkok?”) are answered using the internet search with Perplexity.

To accurately route each user query to the appropriate search mechanism, Vladimir fine-tuned the Mistral 7B Instruct model using our fine-tuning recipe. To make the classification even more accurate, Vladimir further customized the Command Generator.

Another standout feature Vladimir implemented is AI Admin, which activates when a user shows interest in certain topics. It collects user details such as name and email address, leveraging CALM’s business logic component. Once the conversation is complete, the agent automatically sends a summary of the dialogue directly to the user’s email.

A message from Vladimir:

“Big thanks to the RASA team for the 2025 Agent Building Challenge! It's about pushing the limits of what you can do in a time-bound situation with a new dialogue creation tool. ~2 weeks of hard work and one overnight just before the deadline were enough to test CALM in different ways, from rule-based flows to technically sophisticated cases where prompt engineering and LLM fine-tuning techniques were heavily applied. Although I am happy to be the winner of this competition, the most important thing for me is participation and constant readiness to learn new things. I would like to especially thank my wife Anastasiia and family, who gave me the opportunity to give my best!”

🥈2nd Prize - $3,000

Team Eryc, Martin Lukáň, Slovakia – “AI Blood Donation Agent”, GitHub

Watch Martin's demo here:

Martin developed an agent to simplify the blood donation process in Slovakia. He created a multilingual agent that speaks Slovak and English, capable of answering any questions related to blood donation, as well as performing transactional tasks such as filling out a pre-donation questionnaire to assess eligibility, booking an appointment, and modifying or cancelling a visit.

To handle informational queries, Martin implemented an agentic RAG system using FAISS vector search and the GPT-4o-mini model. The data source consisted of information from the official website of the National Transfusion Service.

Transactional skills were implemented across seven flows. The project features complex integrations, including email-based user registration with OTP verification, database integration for managing appointments and storing questionnaire data, and QR code generation for donor identification at the blood donation center.

We were impressed by this project not only for its technical complexity and social relevance but also for its level of personalization: the agent’s avatar is represented as a drop of blood, styled in red colors, and integrated into a web chat.

A message from Martin:

"I really enjoyed participating in the Rasa AI Building Challenge. It was a great opportunity to experiment with new Rasa features, connect with like-minded people, and through the Eryc chatbot - explore the use of AI in healthcare, which has immense potential to improve lives."

🥉3rd Prize - $1,000

Muhammed Emin Tetik, Turkey – “AI Dental Assistant Agent”, GitHub

Watch Muhammed's demo here:

Muhammed developed an AI Dental Assistant Agent to support the work of dental assistants and streamline their daily operations.

The agent includes four main flows:

Registering a new patient
Creating a new appointment
Sending visit reminders
Analyzing medical images

For response generation, Muhammed used the Gemma 3 12B model, running with Ollama on an RTX 3060 GPU. To parse and convert dates and times into datetime format, he integrated Duckling Server.

One of the most impressive parts of the project is the image analysis flow, powered by the Grok-2 Vision model. During the demo, Muhammed presented a jaw X-ray analysis scenario: the user uploads an X-ray image, and the agent responds by asking clarifying questions and then provides a detailed analysis and follow-up recommendations.

A message from Muhhamed:

"Participating in Rasa's AI Agent Building Challenge pushed me to discover the perfect harmony between Rasa’s structured business logic and the expressive capabilities of large language models. I particularly enjoyed having complete freedom in defining my project's theme and approach, which sparked genuine creativity. The supportive attitude from the Rasa team throughout this experience made the entire process motivating and enjoyable. Waiting eagerly for the final results after rushing to meet the video submission deadline added an unforgettable thrill to this adventure. Many thanks to the Rasa team and participants!"

💡Best Feature Request - $1,000

We’ve received a large number of insightful feature requests, with the best feature requested by Kennedy Yinusa, the MailoBot Team, GitHub.

🌟10 Notable Project Winners

#1 Eddyraj Rajiah, United States, GitHub

A dating platform that uses conversational AI instead of swipes to build authentic user profiles based on personality. It matches users through genuine compatibility and even helps plan real-life dates by integrating with restaurant booking services.

#2 KisanTeam, Nepal, GitHub

Kissan agent is designed to help farmers in countries like Nepal buy and sell goods directly, avoiding middlemen. It allows users to list, edit, and purchase agricultural products and even chat with experts for advice, making farming more profitable and connected.

#3 Kryptos Team, India, GitHub

Cryptos is a crypto investment analysis agent designed to give users real-time insights into the crypto market. It uses Mistral LLM and integrates with Telegram.

#4 AI Elderly, Denmark, GitHub

Voice Assistant for elderly people. This assistant can help manage emails. It is connected to Gmail accounts and can read and reply to emails in different styles. It can also create a distress call to an emergency contact.

#5 André Natanael da Silva Freitas, Brazil, GitHub

This AI agent is a specialist in local places. It can recommend cafes, restaurants, beauty salons, hospitals, hotels, pharmacies, and barbershops, and it can also summarize reviews.

#6 Julia, Germany, GitHub

AI Agent that supports young parents. It can summarize videos on baby starter kits, find doctors and book appointments, find and book babysitters, connect to baby tracker apps, and much more.

#7 Two Chatbots, India, GitHub

Revolutionizing property search with Conversational AI. This agent helps with finding a dream home or a perfect rental with the ease of a casual conversation.

#8 Team Godel, India, GitHub

The agent can help build a dream house by customizing every aspect of the user’s choice. Arch bot core is a powerful geometry engine that understands 2D floor plans.

#9 Harivallabha, India, GitHub

Navigation voice agent. This AI agent asks for route information and preferred means of travel. It’s multilingual and can speak two languages.

#10 Asahi Tech India, India, GitHub

SeBrealth is a healthcare AI assistant designed to make women’s health support more accessible and compassionate. It offers a private, judgment-free space to talk about personal topics like breast health, STIs, pregnancy, and reproductive wellness.

What’s Next?

We congratulate the winners and thank all participants for taking part in the Rasa Agent Building Challenge, and invite everyone to stay tuned for our future events. You can also find the full list of winners on our GitHub and watch the replay of the livestream and live demo on Rasa’s YouTube Channel.

If you feel inspired and want to build your own AI Agent using Rasa Pro, install Rasa and follow this simple tutorial.