2025

Building the Next OS-Level AI Experience

A single, powerful multimodal foundation model can serve as an unchangeable firmware within edge/mobile operating system, enabling applications to use compact "adapters" (for varied downstream tasks – text, image, audio, video) instead of bundling several large models. Some of the architectural innovations that can be included are - firmware backbone and task-specific adapters, multi-path execution to route tasks efficiently based on complexity, demonstrating system benefits through metrics like latency and battery performance.

Building the Untethered, Always-On AI Companion

Reimagine a smartphone that doesn't just run apps, but truly understands and assists user. An agent that sees what you see, hears what you hear, and remembers your experiences to provide contextual, real-time help, all without a constant connection to the cloud.

On-Device Multi-Agent System for Behavior-Based Anomaly & Fraud Detection

Multi-agent system that runs fully on-device, continuously learning and modeling user behaviour patterns to detect anomalies or potential fraud in real-time, without sending sensitive data to external servers. The system can monitor user behaviour patterns (e.g., touch patterns, typing rhythm, app usage, movement) and build local models of “normal” behaviour. It should detect and react to anomalous or suspicious activity (e.g., unauthorized access, bot-like behaviour, spoofing).

On-Device Agentic System for Intelligent Battery Optimization

An agentic system that intelligently optimizes battery usage for a target application, operating fully on-device, with no reliance on cloud computation. Adaptive, modular, and context-aware, ensuring the target application continues to function effectively while maximizing battery life.

On-Device Fine-Tuning Framework for Billion+ Parameter scale LLMs

Efficient framework for the on-device fine-tuning of Billion+ scale Large Language Models on a Galaxy S23-S25 equivalent smartphone/edge device. Enable a typical application to adapt a pre-trained LLM to a user's personal data, all while operating within the tight constraints of a mobile environment.

Crafting the Next Generation of Human-AI Interaction on the Edge

Let's go beyond model and traditional interaction capabilities on any edge device. A solution that addresses a real-world problem by leveraging on-device Generative AI, while pioneering novel, effective, and intuitive Human-AI Interaction (H-AI).

Estimation of UAV Parameters Using Monostatic Sensing in ISAC Scenario

Develop an AI-based solution using monostatic integrated sensing and communication (ISAC) to estimate UAV range, velocity, and direction of arrival, leveraging advanced signal processing and machine learning. Utilize the channel model based on 3GPP TR 36.901-j00 (Rel-19) Section 7.9 for ISAC applications. Participants are expected to design models that extract these parameters from ISAC signals under the specified channel conditions.

Classify User Application Traffic at the Network in a Multi-UE Connected Scenario

Applications are affected differently under varying traffic conditions, channel states, and coverage scenarios. If the traffic of each UE can be categorized into broader categories, such as Video Streaming, Audio Calls, Video Calls, Gaming, Video Uploads, browsing, texting etc. that can enable the Network to serve a differentiated and curated QoS for each type of traffic. Develop an AI model to analyze a traffic pattern and predict the application category with high accuracy.

Real-time Detection of Reel Traffic vs Non-reel Traffic in a Social-networking Application

SNS applications (such as Facebook and YouTube), transmit both video (short videos, reels, etc.) and non-video traffic (feeds, suggestions, etc.) through the same data pipeline. Develop an AI model to differentiate reel / video traffic versus non-reel/video traffic in real-time, enabling user equipment (UE) to optimize performance dynamically. The model should also ensure accuracy under varying network congestion and coverage conditions.

AI challenge will be conducted in 2 phases:

• Phase 1 (Online): All participating teams will work on a problem statement and submit their solution as per the submission guidelines before the deadline.
• Phase 2 (Bengaluru, On-site Finals): AI Challenge finalists (top 5 teams selected based on the Phase 1 evaluation) will present their solution in the Samsung EnnovateX 2025 event in Bengaluru and compete for the Grand Prize.

• This contest is open to all individuals above the age of 18 and currently residing in India.
• Samsung employees cannot participate in this contest.
• Participation can be solo or in a team of upto 4 members.
• Team leader must fill the registration form with all member details after finalizing the problem statement.
• Each person can be a part of only 1 team.
• If the team is selected for finals, all travel and accommodation arrangements & expenses has to be borne by the team.
• Identity and Age verification will be performed for AI Challenge finalists (all team members).

General Guidelines

• Your submission must be capable of being successfully installed/executed and must run consistently on the intended platform.
• Your submission must function as demonstrated in the demo video and/or expressed in the technical documentation.
• Your submission must be a new project. If the project existed prior to the AI Challenge contest period, it should be significantly updated during the contest period. You must explain how the project has been significantly updated during the contest period in the documentation.
• Your submission must abide by the "Open Source Rules & Disclosure".

Submission Requirements

• Your AI challenge submission must be curated in a public GitHub repository.
• Before the contest deadline you must e-mail your repository link to ennovatex.io@samsung.com via the team leader's e-mail id provided at the time of the registration. Apart from member names, problem statement & your GitHub repository submission link, no other project information must be provided in the email. All project information should be documented in the GitHub repository.
• The repository must be public and must have one of the following open source licenses - MIT, Apache 2.0 or BSD 3-Clause.

Permitted Resources

Participants may use:

• Open-source libraries (e.g., licensed under OSI-approved licenses like MIT, Apache 2.0, GPL).
• Open data (publicly available datasets under licenses like Creative Commons, Open Data Commons, or equivalent).

Participants are allowed to publish(open source) any synthetic or proprietary dataset used in their project, but will be responsible for any legal compliance and permission for the same. The dataset can be published under Creative Commons, Open Data Commons, or equivalent license.

Prohibited Resources

Participants must not use:

• Third-party APIs (e.g., commercial APIs, cloud services, or any external API requiring authentication or payment).
• Third-party SDKs (e.g., proprietary software development kits or frameworks).
• Proprietary data (e.g., non-public datasets, licensed/internal data, or data requiring restricted access)

Phase 1 (Online)

The evaluation criteria for Phase 1 are provided below:

• Novelty of Approach - 25%
• Technical implementation & Documentation - 25%
• UI/UX Design or User Interaction Design- 15%
• Ethical Considerations & Scalability - 10%
• Demo Video (10 mins max) - 25%

Phase 2 (Bengaluru, On-site Finals)

The evaluation criteria for Phase 2 are provided below:

• Unique Value Proposition - 30%
• Presentation & Flow - 30%
• Ethical Considerations & Scalability - 10%
• Demonstration - 30%

• 1 Grand Prize - INR 1 Lakh
• 1 Runner-up Prize - INR 50 Thousand

• Harassment, discrimination, or abuse of any kind will not be tolerated.
• All participants must adhere to a respectful and inclusive environment.
• Plagiarism or use of unlicensed IP will result in disqualification.
• AI-generated code must be attributed and clearly explained.

For any query or support, please feel free to reach out to us at ennovatex.io@samsung.com.