From Research to Reality: How Everpure Is Delivering on the Key AI Trends of NeurIPS 2025

This post was co-authored by Gauri Kholkar, Priyanka Choudhary, Ruchi Awasthi (Applied AI/ML Scientist, Office of the CTO), and Dr. Ratinder Paul Singh Ahuja (CTO for Security and GenAI). Dr. Ahuja is a renowned name in the field of security, AI, and networking.

Researchers today are asking one critical question: “How can we make AI run faster, act autonomously, and stay secure?”

At Everpure, we see a direct line between these research themes and the challenges our enterprise customers face every day. Building on our momentum from the 2025 ACL Conference, we were thrilled to make our debut this year at the 2025 Conference on Neural Information Processing Systems (NeurIPS), the premier global event for machine learning. 

This year, the message from the world’s top researchers was clear: The “brute force” era of AI is evolving. While scaling laws still matter, the focus has shifted to efficiency, agency, and safety. We were proud to contribute directly to this prestigious dialogue with multiple accepted papers and presentations. Read on to discover the latest projects we shared that align with the biggest breakthroughs at NeurIPS 2025.

1. The Era of Agentic AI and Enterprise Copilots

At NeurIPS, the rallying cry was “From Seeing to Doing.” At Everpure, we’re moving beyond simple chatbots to true agentic workflows. We’re deploying specialized enterprise agents designed for high-stakes domains like knowledge management and finance, and bringing manageability to the data center with the Everpure AI Copilot. 

AI Copilot combines the power of our storage dashboards with the simplicity of natural language, allowing admins to ask questions like “Why did latency spike at 2 am?” and receive instant, actionable insights.

Everpure researchers Priyanka Choudhary and Ruchi Awasthi presenting their work on agentic AI at NeurIPS.

We showcased our research in this area with two key papers presented at the Women in Machine Learning (WiML) Workshop:

• “Hybrid AI for Automating High-Stakes Procurement Workflows” by Priyanka Choudhary and Ratinder Ahuja (selected among top 15% of accepted papers).
  This work addresses the operational bottlenecks in purchase requisition (PR) approvals using a multi-agent architecture (LLM-EXTRACT, LLM-VAL, and LLM-REPORT-GEN). By integrating deterministic business rules with a hybrid confidence score, the system achieved a 62% reduction in manual review time, a 37% faster procurement cycle, and a 66% decrease in submission errors across 4,000 annual requests.
• “Mitigating Knowledge Entropy: A Multi-Agent Framework with Decoupled Re-ranking & Governance-by-Design” by Ruchi Awasthi and Gauri Kholkar.
  To combat “knowledge entropy” in enterprise databases, this paper proposes a framework featuring context-aware reranking (CAR) for handling ambiguity and a multi-step actionable verdict engine (AVE). Crucially, it incorporates a dynamic AI reliability score to prioritize human-in-the-loop oversight. In production, this system improved verdict accuracy to 88.2%, reduced authoring time by 75%, and yielded an estimated $430k in annualized savings.

Other research focused on enabling agents to execute complex workflows reliably:

• Financial precision: Papers like “Instruction Following for Finance” and “FISCAL” demonstrate how agents are evolving to verify financial documents and handle regulatory rules with high precision.
• Reliability bottlenecks: Scaling these systems isn’t trivial. Research such as “Why Do Multi-Agent LLM Systems Fail?” highlights critical failure points like context loss and inter-agent misalignment that must be solved for enterprise adoption.
• Human-computer interaction: Frameworks like MAC are validating the use of LLMs for complex, interactive data tasks and clarifying user ambiguity—key requirements for tools like AI Copilot.

2. Safety by Design: Automated Guardrails

With great power comes great responsibility. Another focus this year was on sophisticated detection and human-centric safety. 

At Everpure, we believe security should be built-in, not bolted-on. Our leadership here was recognized with an Oral Presentation at the Regulatable ML Workshop:

Gauri Pradipkumar Kholkar delivering her spotlight oral presentation on “policy-as-prompt” at NeurIPS.

• “Policy-as-Prompt: Turning AI Governance Rules into Guardrails for AI Agents” (Spotlight talk) by Gauri Pradipkumar Kholkar and Ratinder Ahuja.
  This work introduces a regulatory ML framework that automatically converts unstructured design artifacts—like PRDs, TDDs, and code—into verifiable runtime guardrails. By building a source-linked policy tree, the system compiles lightweight classifiers that enforce least privilege and data minimization in real time. This “policy-as-code” approach not only reduces prompt injection risks but also provides complete traceability and audit logging, bridging the gap between static governance policies and dynamic AI agent behavior.

Other research to explore:

• Consortium verification: “Teaming LLMs to Detect and Mitigate Hallucinations” proposes using “teams” of models to cross-verify outputs and reduce errors.
• Human-centric bias: “Sycophancy Claims about Language Models” emphasizes evaluating how models might bias outputs to please users.
• Auditing: Research like MELISSA highlights the need for agents that can scrutinize their own outputs and maintain transparent audit trails.

3. AI for Security and IT Operations (AIOps)

The focus on security extended beyond model weights to the infrastructure itself. New research highlights how AI is transforming IT operations and threat detection. 

At Everpure, we’re building resilience directly into the infrastructure layer. Today’s security operations centers are overwhelmed. We advocate that SOCs must embrace AI as a practical, scalable path to reducing alert fatigue, false positives, and burnout. 

While AI can help, it’s not as simple as plug and play; it requires intelligent integration. We leverage AIOps within our platform to predict workloads and detect anomalies before they become outages. Furthermore, we provide rapid recovery from cyber threats through immutable snapshots, ensuring that data remains secure even when perimeter defenses are breached.

Other research to explore:

• Automated root cause analysis: “ALICE: Agentic Logic for Incident and Codebug Elimination” demonstrates a multi-agent system that links observability data (logs, traces) with source code to pinpoint the root cause of production incidents, moving beyond simple symptom detection.
• Real-time intrusion detection: GraphFaaS introduces serverless graph neural networks to detect intrusions in real time, handling the “bursty” nature of cyber threats efficiently.

4. The AI Developer: Generating Workflows from Scratch

The message was clear: Generative code is moving from “completion” to “end-to-end engineering.” At Everpure, we’re also focused on pushing the boundaries of developer productivity. We’re building systems that generate entire coding workflows directly from requirements. This creates a “software factory” where AI scaffolds necessary logic, tests, and data pipelines, bridging the gap between a product manager’s vision and an engineer’s implementation.

Research from NeurIPS to dive into:

• Self-correction: “Lessons Learned” explores how coding agents can self-correct and improve over time.
• Rigorous benchmarks: “Auto-SWE-Bench: Scalable, Real-World Benchmarks for LLM Coding Evaluation” establishes new standards for evaluating these autonomous builders on complex, real-world tasks.

5. Smashing the Inference Bottleneck: The Memory Wall

At NeurIPS, the “memory wall” was a dominant theme—specifically the bottleneck created by the key-value (KV) cache. At Everpure, we’re tackling the memory wall head-on with a storage-first approach to inference, developing a solution to offload the KV cache from expensive GPU memory to high-performance storage via NFS and S3. 

We leverage the low latency of FlashBlade® to achieve up to 20X faster inference compared to standard implementations. This innovation effectively “swaps” memory to flash, allowing enterprises to run massive context windows without the prohibitive cost of excessive GPU infrastructure.

Figure 3. Performance of Pure KVA on NFS and S3 vs. vLLM on Llama 3.1 70B.

New research validates our storage-centric approach:

• Swapping to flash: “HiFC: High-efficiency Flash-based KV Cache Swapping” introduces a novel DRAM-free scheme leveraging direct access to high-speed SSDs to sustain throughput.
• Dynamic precision: QAQ proposes dynamic precision to reduce memory footprints without sacrificing accuracy.
• Massive scale: Papers like PORT offer new reinforcement learning techniques for distributing models across thousands of GPUs and efficiently routing queries.

Conclusion

NeurIPS 2025 showed us the future of AI: efficient, agentic, safe, and collaborative. At Everpure, we’re building the platform that powers that future today. Whether it’s offloading KV caches to FlashBlade to break inference limits or deploying safe, autonomous agents for finance, we’re ensuring that your data infrastructure is as advanced as the models running on it. 

Learn more about the world’s most powerful data storage platform for AI. Check out our Pure360 series on powering AI at scale.