Research Paper

Why It Matters

Edge devices currently lack versatile on-device learning, limiting personalization and adaptability. This solution reduces latency, energy use, and privacy risks by enabling multiple learning modes locally. It transforms workflows by allowing smart devices to continuously learn and adapt in real time, scaling across industries like healthcare and consumer electronics.

Potential Customers & Pain Points

• Smart device manufacturers – Need real-time personalization without cloud
• Healthcare providers – Require adaptive patient monitoring
• IoT platform developers – Seek energy-efficient on-device learning
• Consumer electronics brands – Demand privacy-preserving AI features

Research Paper

Why It Matters

Tokenization overhead dominates latency in LLM serving, especially for agentic systems that append long transcripts frequently. TokTier reduces redundant tokenization work, enabling faster response times and higher throughput. This efficiency gain lowers infrastructure costs and improves user experience, making large-scale agentic LLM deployments more practical and scalable.

Potential Customers & Pain Points

• AI platform providers – High tokenization latency limits throughput
• Cloud service operators – High compute cost for repeated tokenization
• Enterprises deploying agentic LLMs – Inefficient session state reuse slows workflows
• Developers of LLM-based tools – Long response times degrade user experience

Research Paper

Why It Matters

Large language models require significant computational resources, making energy-efficient inference critical for deployment at scale. LightRot reduces energy consumption while maintaining accuracy on advanced models, enabling cost-effective and sustainable AI services. This efficiency supports broader adoption in industries relying on conversational AI and large-scale language processing.

Potential Customers & Pain Points

• Cloud providers – High inference energy costs
• AI service developers – Need accurate low-bit model deployment
• Edge device manufacturers – Limited power and compute resources

Research Paper

Why It Matters

As AI models grow larger and more complex, efficient hardware is critical to manage power and performance. SpiNNaker2 addresses this by combining neuromorphic and deep learning capabilities on a single platform, enabling flexible, scalable, and energy-efficient computation. This supports diverse AI applications and accelerates innovation in brain-inspired computing at scale.

Potential Customers & Pain Points

• AI hardware developers – Need energy-efficient scalable AI chips
• Neuromorphic researchers – Require flexible platforms for spiking neural networks
• Edge device manufacturers – Demand low-power high-performance AI processing
• Data centers – Seek cost-effective acceleration for deep learning workloads

Research Paper

Why It Matters

Video communication under low bandwidth and unstable networks often suffers from poor visual quality and inefficiency. This solution improves transmission efficiency and robustness while maintaining perceptual quality, enabling reliable video services in constrained environments. It transforms workflows by reducing data needs and enhancing user experience in remote, mobile, and emerging markets.

Potential Customers & Pain Points

• Telecom operators – Need efficient video delivery over weak networks
• Streaming platforms – Need to reduce bandwidth costs while preserving quality
• Remote work and education providers – Need reliable video under unstable connections
• IoT and surveillance systems – Need low-bandwidth video transmission with high perceptual utility

Research Paper

Why It Matters

Training large language models requires massive GPU clusters running for months, where hardware failures cause costly delays and resource waste. DeadPool reduces downtime and eliminates checkpoint overhead, enabling more efficient and resilient training workflows. This improves productivity and lowers operational risks for AI research and enterprises scaling LLM development.

Potential Customers & Pain Points

• AI research labs – Need to minimize training interruptions
• Cloud GPU providers – Need to improve resource utilization and reduce failure impact
• Enterprises training LLMs – Need cost-effective fault tolerance for large-scale models

Research Paper

Why It Matters

As privacy, latency, and cloud cost concerns push AI inference to edge devices, BaseRT enables high-performance local LLM execution on Apple Silicon. This reduces reliance on cloud infrastructure, lowers operational costs, and improves user experience by delivering faster responses. It supports a broad range of models and quantisation formats, making it scalable across device generations and application needs.

Potential Customers & Pain Points

• AI app developers – Need efficient on-device LLM inference
• Enterprises – Require privacy-preserving AI with low latency
• Cloud providers – Seek to reduce inference costs
• Hardware OEMs – Want optimized software for Apple Silicon capabilities

Research Paper

Why It Matters

Uplink-dominant 6G applications like cooperative vehicular streaming face bandwidth constraints transmitting large visual data. Reducing redundant transmissions improves network efficiency and user experience, enabling scalable, high-fidelity data sharing in dense urban environments. This approach supports sustainable and resource-efficient 6G uplink systems critical for future connected mobility.

Potential Customers & Pain Points

• Telecom operators – Need to optimize uplink bandwidth
• Automotive OEMs – Require reliable cooperative vehicular data sharing
• Smart city planners – Demand scalable urban connectivity solutions
• 6G infrastructure providers – Seek efficient resource management.

Research Paper

Why It Matters

Training large and heterogeneous AI models requires efficient optimizers that scale without excessive computational overhead. DMuon reduces optimizer latency significantly, enabling faster training cycles and cost savings. This efficiency gain supports scaling complex models and accelerates AI innovation workflows.

Potential Customers & Pain Points

• AI research labs – High training costs and slow optimizer steps
• Cloud AI service providers – Need scalable efficient distributed training
• Enterprises developing large language models – Require faster model iteration and deployment.

Research Paper

Why It Matters

LLM API costs and latency are major bottlenecks for enterprises deploying AI at scale. RLM-Cascade reduces these costs by nearly half and speeds up response times, enabling more efficient and cost-effective AI services. This approach scales across diverse workloads without requiring model internals, making it practical for broad industry adoption.

Potential Customers & Pain Points

• Enterprises using LLM APIs – High inference costs
• AI service providers – Latency and throughput constraints
• Cloud platform operators – Resource inefficiency
• Software developers – Need for reliable fast AI coding assistance

Research Paper

Why It Matters

LLM training efficiency is limited by batch construction blind to true sample costs, causing wasted GPU resources and slower training. ODB addresses this by dynamically batching with accurate cost awareness, improving throughput and reducing padding overhead. This transforms fine-tuning workflows by enabling faster, scalable training on heterogeneous data without costly preprocessing or kernel modifications.

Potential Customers & Pain Points

• AI research labs – Inefficient LLM training throughput
• Cloud ML platforms – High GPU resource waste
• Enterprises fine-tuning LLMs – Slow and costly model updates
• AI infrastructure providers – Need scalable compatible batching solutions.

Research Paper

Why It Matters

LLM training efficiency is limited by blind batch formation that ignores true sample costs, causing wasted GPU resources and slower training. ODB improves throughput significantly while maintaining model quality and synchronization, reducing costs and accelerating development cycles. This scalable approach benefits enterprises fine-tuning large models on diverse datasets without complex infrastructure changes.

Potential Customers & Pain Points

• AI research labs – Inefficient LLM training throughput
• Cloud ML platforms – High GPU costs from padding and memory waste
• Enterprises fine-tuning LLMs – Need scalable cost-effective batch processing
• ML infrastructure providers – Demand for drop-in compatible batching solutions.

Research Paper

Why It Matters

Training extremely large language models typically requires massive distributed hardware, limiting access and increasing costs. This approach reduces hardware barriers by enabling end-to-end training of hundred-billion-parameter sparse models on a single node, lowering costs and accelerating experimentation. It democratizes large-scale model development for research labs and enterprises with limited infrastructure.

Potential Customers & Pain Points

• AI research labs – High cost and complexity of large model training
• Cloud providers – Need to optimize resource usage for large model workloads
• Enterprises – Limited access to large-scale AI due to hardware constraints
• AI startups – Need scalable cost-effective training solutions.

Research Paper

Why It Matters

3D video streaming demands high bandwidth and computational resources due to large frame sizes and complex scene representations. GS-NFS reduces encoding and decoding latency drastically, enabling real-time streaming of dynamic 3D content with competitive quality. This efficiency supports scalable deployment in applications like VR, AR, and remote collaboration, transforming workflows by making high-quality 3D video practical over variable network conditions.

Potential Customers & Pain Points

• VR/AR platform providers – Need real-time high-quality 3D streaming
• Cloud gaming companies – Require low-latency 3D content delivery
• Remote collaboration tools – Demand bandwidth-efficient dynamic 3D video
• 3D content creators – Face slow compression workflows limiting iteration speed

Research Paper

Why It Matters

5G URLLC applications require ultra-low latency and high reliability, but current uplink scheduling incurs significant delays and resource waste. AUGUSTE reduces round-trip latency to meet stringent 5G targets while drastically lowering resource consumption, enabling scalable, efficient support for industrial automation, V2X, and edge control systems. This improves network responsiveness and cost-efficiency for critical real-time services.

Potential Customers & Pain Points

• Telecom operators – Need to meet URLLC latency SLAs efficiently
• Industrial automation firms – Require reliable low-latency wireless control
• Automotive OEMs and V2X providers – Need ultra-responsive vehicle communication
• Edge computing providers – Demand optimized uplink scheduling for real-time inference.

Research Paper

Why It Matters

As AI demand grows, power grids face capacity and cost challenges. Deploying AI compute at renewable sites creates local demand, reduces transmission losses, and eases grid strain. This approach enables scalable, sustainable AI infrastructure expansion aligned with renewable energy availability.

Potential Customers & Pain Points

• Cloud providers – Need to reduce inference latency and power costs
• Renewable energy operators – Need to monetize excess capacity
• AI service providers – Need scalable sustainable compute infrastructure
• Utilities – Need to balance grid load and integrate renewables.

Research Paper

Why It Matters

Mental health disorders affect millions and overwhelm healthcare systems with vast clinical data. This framework automates and stabilizes mental health screening at population scale, improving efficiency and trustworthiness. It supports healthcare providers in managing large datasets while adapting to patient-specific needs, enabling broader and more consistent mental health monitoring.

Potential Customers & Pain Points

• Healthcare providers – Overwhelmed by clinical data volume
• Mental health organizations – Need scalable screening tools
• Telemedicine platforms – Require adaptive patient-specific analysis
• Public health agencies – Demand population-level mental health insights.

Research Paper

Why It Matters

AI inference is a growing electricity demand source with flexibility to shift computation geographically. Efficient relocation reduces energy costs and carbon emissions while respecting latency and regulatory limits, enabling scalable, sustainable AI services globally.

Potential Customers & Pain Points

• Cloud providers – High energy costs and carbon footprint
• Data center operators – Capacity and regulatory constraints
• Enterprises with AI workloads – Need latency-compliant cost-efficient inference
• Sustainability-focused tech firms – Demand carbon-efficient AI operations

Research Paper

Why It Matters

Enterprise AI applications increasingly rely on complex compound AI systems that require efficient, scalable inference infrastructure. This architecture reduces latency and cost while supporting concurrent multi-model workflows, enabling faster deployment and iteration of AI agents at scale. It transforms AI operations by handling bursty workloads and heterogeneous scaling, critical for real-world enterprise adoption.

Potential Customers & Pain Points

• Enterprises deploying AI agents – Need scalable low-latency inference
• Cloud service providers – Need cost-effective multi-model serving
• AI platform developers – Need support for rapid model iteration and bursty workloads

Research Paper

Why It Matters

Large language model serving faces critical memory bottlenecks due to KV-cache size, impacting latency and throughput. This solution reduces memory footprint without sacrificing accuracy or system compatibility, enabling scalable, cost-effective deployment of LLMs in production environments. It improves efficiency for providers handling diverse workloads and concurrency levels.

Potential Customers & Pain Points

• Cloud providers – High memory costs and latency in LLM serving
• AI service platforms – Need scalable efficient LLM deployment
• Enterprises deploying LLMs – Limited hardware resources and throughput constraints