Case Studies & Solutions

Elevating Last Mile Autonomous Delivery with “Last-inch” intelligence

We started with something that was already pretty remarkable - smart mailboxes that could accept deliveries from drones and robots, a first in the delivery industry. These units were doing what no other mailbox could: creating secure drop-off points for autonomous deliveries in residential areas. But we saw a chance to make them even better.

Our team added artificial intelligence to these mailboxes, essentially giving them the ability to think, see, hear, and make decisions. Instead of just being really good at receiving packages, they became active participants in the delivery process. We equipped them with sensors and AI models that could process what was happening around them in real-time.

The system works on two levels. First, each mailbox can now think for itself - making decisions about packages and interacting with delivery vehicles more intelligently. Then there's the bigger picture: all these smart mailboxes talk to our cloud platform, sharing what they learn and helping us spot patterns across the entire network. This means we can predict busy periods, optimize delivery schedules, and keep improving how the whole system works.

This combination of innovative hardware and AI has opened up new possibilities. The mailboxes aren't just receiving packages anymore - they're helping make the entire delivery process smarter. They can coordinate better with drones and robots, track package conditions, and give us insights we never had before.

Looking forward, we've built this system so it can grow and handle new challenges. As cities get smarter and delivery methods evolve, these intelligent mailboxes are ready to adapt and take on more complex tasks. What started as a breakthrough in package delivery has become even more valuable with the addition of AI, showing how we can take great ideas and make them even better through intelligent automation.

Making Data Work for Everyone: A major Telco’s Journey to Intelligent Data Discovery

At a major US based Telco, data is everywhere - but finding the right information wasn't always easy. Picture about 60,000 employees, each needing to make quick, informed decisions, but having to navigate through a maze of legacy systems and thousands of data tables spread across different platforms. It was like having a massive library without a proper catalog system - the information was there, but finding it was a real challenge.

For the Chief Data Office (CDO), this wasn't just about organizing data; it was about transforming how their quarter-million employees interact with information daily. Many employees were spending valuable time jumping between different applications, trying to piece together the data they needed. Some were using systems that hadn't changed much in years, while others had developed their own workarounds just to get their jobs done.

We approached this challenge by asking a simple question: What if accessing company data was as easy as using a modern search engine? This led us to develop a central data platform that brings together all of the Telco’s data resources in one place. But we didn't stop at just organizing the data - we added artificial intelligence to make the system truly smart.

The platform we built works like a highly intelligent librarian and an information Genie on steroids. Using generative AI, it helps employees find exactly what they need, even if they're not sure about the technical names or locations of the data. The system understands natural language queries, connects related datasets, and even suggests relevant information that users might not have known to look for.

Some key features we implemented include:

• Smart search that understands business context and user intent

• Automated data quality checks to ensure information reliability

• Natural language querying that lets employees ask questions in plain English

• AI-powered recommendations that suggest related datasets

• A unified interface that brings together data from various sources

The impact has been significant. What used to take hours of searching across multiple systems can now be done in minutes. Employees who weren't data experts can now confidently find and use the information they need. The system doesn't just store data - it makes data truly accessible and useful for everyone, from marketing teams looking at customer trends to operations teams optimizing network performance.

Looking ahead, this platform is designed to grow smarter with use. As more employees interact with it, the AI learns and improves its recommendations and search results. It's not just solving today's data challenges - it's building a foundation for how Telco's workforce will work with data in the future.

This transformation shows how the right combination of organization and intelligence can turn overwhelming amounts of data into a valuable, accessible resource for everyone in the company. It's about more than just finding data - it's about empowering every employee to make better, data-driven decisions.

Closing the Gap with Real-Time Voice Agents

Traditional voice AI systems have struggled with high latency and unnatural interactions, creating a significant barrier to enterprise adoption. While these systems could handle basic queries, the noticeable delays and poor conversational flow made them unsuitable for complex business processes that require real-time interaction and decision-making.

We developed a solution that fundamentally transforms voice AI capabilities through a sophisticated multi-agent architecture. At its core, we implemented WebRTC-powered voice streaming using LiveKit and Ultravox platforms, achieving sub-120ms latency - a critical breakthrough that enables truly natural conversation flows.

Our multi-agent framework orchestrates several key components working in parallel:

• Real-time speech-to-speech generation with near-zero latency

• Continuous context management and memory systems

• LLM-based reasoning for intelligent response generation

• Direct task execution through function calling

• High-fidelity speech-to-text and text-to-speech synthesis

The system's architecture is designed for enterprise-grade performance, featuring high availability and robust security measures. By leveraging peer-to-peer connections, we've created a scalable solution that significantly reduces infrastructure costs while maintaining consistent performance under load.

This technology has been successfully deployed across multiple enterprise environments, demonstrating its versatility and effectiveness. The integration of agentic frameworks with reasoning models has enabled these voice agents to handle complex, context-dependent tasks while maintaining natural conversation flows - something previously unattainable with traditional voice AI systems.

The impact extends beyond just improved customer interactions. Organizations using this solution have gained the ability to automate sophisticated business processes while maintaining the quality of human-like interactions. As language models and voice technology continue to advance, this framework provides a foundation for even more capable voice-based AI interactions in enterprise settings.

Shaping AI Code Assistance: Expert-Driven RLHF Evolution

When one of the leading technology companies invited us to help improve their code-assist LLM, we faced a unique challenge. The goal was to enhance their model using Reinforcement Learning from Human Feedback (RLHF), but the field was still experimental with no established best practices for creating effective RLHF training data for code assistance.

We assembled a global team of subject matter experts to create high-quality feedback data for the RLHF process. These weren't just programmers - they were experienced developers who could effectively judge model outputs and provide nuanced feedback about what constitutes good code suggestions. Our team grew to include several tens of experts, each bringing unique perspectives on code quality, style, and best practices.

The heart of our approach was creating a "golden dataset" for RLHF training - combining expert-written examples with real developer queries. This dual-source approach was crucial: experts provided high-quality preferences and feedback data, while real user queries ensured the model was trained on authentic developer needs. The feedback data helped the model learn which responses were more helpful, more accurate, and better aligned with professional coding standards.

Working closely with the client's AI team, we established an iterative refinement process:

• Expert team provided comparative feedback on model outputs

• Integration of feedback from real-world developer interactions

• Model fine-tuning using RLHF

• Performance evaluation and analysis

• Dataset refinement based on model behavior

• Repeat with adjusted focus areas

Each iteration brought new insights about effective RLHF data creation for code assistance, allowing us to continuously refine our approach. The experimental nature of the project required us to be highly adaptable, quickly adjusting our feedback strategies based on how the model learned from human preferences.

This pioneering work helped establish foundational practices for RLHF-based fine-tuning of code-assist models and demonstrated the critical role of expert human feedback in improving AI code assistance capabilities.

Knowledge at Speed: Transforming Oil & Gas Operations with AI

In the complex world of midstream oil and gas operations, every minute of downtime counts. A mid-sized US company operating multiple refineries and exchanges faced a critical challenge: helping both new and experienced technicians quickly access and apply the vast knowledge accumulated over years of operations.

Traditional approaches to knowledge management weren't cutting it. New employees spent months learning processes and technical aspects, while experienced technicians often spent hours searching through documentation to troubleshoot issues. Each problem typically presented a unique combination of challenges, making quick resolution difficult even with years of experience.

We developed a solution that puts decades of operational knowledge at technicians' fingertips. Built on AWS infrastructure, the system combines semantic search with generative AI to make complex technical information instantly accessible. Whether in the control room or in the field, technicians can now query the system using natural language - even by voice - and receive relevant, contextual guidance within seconds.

The system works across multiple layers:

• Intelligent document processing that understands technical content from diverse sources

• Semantic search powered by Amazon Kendra that grasps the intent behind queries

• Retrieval-augmented generation using Claude to provide precise, contextual answers

• A mobile-friendly interface that works seamlessly in field conditions

What makes this solution particularly effective is its ability to understand the relationships between different technical concepts. Using ontology-based tagging and sophisticated ranking mechanisms, it can connect seemingly unrelated issues that share underlying causes. This means technicians don't just get documented solutions - they get insights from similar cases across different facilities.

Security was paramount given the sensitive nature of operational data. We implemented multiple layers of protection, including private subnets for data isolation, encryption at rest, and strict access controls integrated with existing authentication systems.

The impact has been significant. New employees now ramp up faster with AI-assisted learning, while experienced technicians can resolve complex issues more quickly. The system continues to learn from each interaction, making the knowledge base more valuable over time. It's not just about storing information - it's about making every bit of operational experience instantly actionable.

Training Robotic Arms in the Simulated World

Handling sensitive packages with robotic arms presents a unique challenge - every package is different, and mistakes can be costly. When a client approached us with their patented automated package handling system, they needed their robotic arms to learn adaptive handling without risking damage to real packages during the training process.

Moving beyond traditional preprogrammed movements, we developed a comprehensive virtual training pipeline using NVIDIA's simulation stack. The solution combined Isaac Sim, Metropolis, and Omniverse to create physics-accurate virtual environments using Universal Scene Description (USD). This enabled us to simulate thousands of package handling scenarios simultaneously.

Our end-to-end workflow encompassed three key phases:

• Data processing that combined synthetic data with real robot interactions

• Policy training using Isaac Lab's reinforcement learning framework

• Validation in Isaac Sim before deployment to physical robots

The power of NVIDIA's framework came from its ability to leverage vast datasets and pretrained knowledge to scale the learning process. Using the OSMO platform, we orchestrated complex training workflows across different compute environments, simulating how the robotic arms would handle packages of varying sizes, weights, and fragility levels. This parallel learning approach dramatically accelerated the development of robust handling policies.

Key aspects of our solution included:

• Physics-accurate simulation of package-arm interactions

• Reinforcement learning for optimizing grip strength and movement

• Parallel training across multiple scenarios

• Deployment of trained policies to NVIDIA Jetson computers on physical robots

The result was a robust system that could adapt to different package types while maintaining careful handling. The robotic arms learned to make real-time decisions based on package characteristics, significantly reducing the risk of damage during handling. This virtual-first approach to robotic training not only reduced development time and costs but also enabled continuous improvement without operational disruption.

Looking ahead, this framework provides a foundation for expanding the capabilities of robotic systems, allowing them to learn and adapt to new handling challenges while maintaining high safety and reliability standards.