I’m thrilled to sit down with Dominic Jainy, a seasoned IT professional whose deep expertise in artificial intelligence, machine learning, and blockchain has made him a thought leader in applying cutting-edge technologies across industries. With a keen eye on the challenges and opportunities in AI deployment, Dominic offers invaluable insights into why many AI agent projects struggle to move from concept to production and how real-time data and innovative architectures can bridge that gap. In our conversation, we explore the hurdles of scaling AI systems, the importance of real-time data access, the role of event-driven designs, and strategies to ensure trust and reliability in AI applications.
Can you walk us through the biggest hurdles you’ve noticed when AI projects try to transition from proof-of-concept to real-world implementation?
Absolutely. One of the primary challenges is that many AI projects look promising in a controlled, prototype environment but falter when exposed to the complexities of real-world data and systems. Organizations often underestimate the integration effort needed to connect AI agents with existing tools and databases. This leads to a disconnect where the AI can’t access the real-time context it needs to make relevant decisions. Scalability is another huge issue—prototypes are often built with small datasets or limited user interactions, but in production, the volume and variability of data can overwhelm the system. I’d say a significant percentage of projects—probably over half—get stuck at this stage, unable to move forward without major reengineering.
What causes multi-agent workflows to become what you’ve called ‘brittle monoliths,’ and how does that affect their ability to scale?
When we talk about multi-agent workflows becoming brittle monoliths, we’re referring to systems where multiple AI agents are tightly coupled together without flexible integration points. Initially, they might seem modular, but over time, as dependencies grow, any change or failure in one part can cascade and break the entire system. This rigidity makes scaling a nightmare because you can’t easily add new agents or adapt to higher loads without risking instability. It impacts reliability too—if one agent struggles with real-time data or a specific task, the whole workflow can grind to a halt, eroding trust in the system.
Why do you believe access to real-time data is so essential for AI agents to perform effectively in production environments?
Real-time data is the lifeblood of effective AI agents because it provides the immediate context needed for relevant and timely decisions. Without it, agents are working with outdated or incomplete information, which can lead to poor outcomes. Take online retail as an example—customers expect hyper-personalized experiences, like tailored product recommendations the moment they browse a site. If the AI doesn’t have real-time access to their browsing history or inventory updates, it can’t deliver that. Merging transaction processing with analytics in real time shortens decision-making cycles, allowing businesses to react instantly rather than waiting for end-of-day reports.
How does adopting an event-driven architecture enhance the visibility and performance of AI agents?
Event-driven architecture is a game-changer because it allows AI agents to react to specific business events as they happen, rather than polling for updates or working off static data. This setup mirrors lessons from microservices, where systems are designed to be loosely coupled and responsive to triggers. For large language models, this means better tracking of what events or inputs they’re responding to, which boosts transparency. It also improves performance since agents aren’t bogged down by unnecessary processing—they act only when relevant events occur, making the system more efficient and easier to monitor.
Can you explain the process of grounding large language models with specific questions and domain-specific data to reduce issues like hallucinations?
Certainly. Grounding large language models, or LLMs, involves anchoring their responses in verified, domain-specific data and framing inputs as precise questions to limit ambiguity. This helps prevent hallucinations—those times when models generate plausible but incorrect information. Techniques like retrieval-augmented generation, or RAG, play a big role here by pulling in relevant data from trusted sources to inform the model’s output. I see this as more of a data and trust issue than a flaw in the model itself. If you don’t feed the model high-quality, contextual data, or if users can’t trust the outputs, no amount of model tweaking will fully solve the problem.
How do streaming agents, as a concept, address some of the production challenges faced by AI deployments?
Streaming agents tackle production challenges by enabling real-time monitoring and action on business events. They operate on platforms that process data streams continuously, often with state management to keep track of context over time. This means they can orchestrate intelligent automation instantly—think of an agent detecting a sudden spike in website traffic and adjusting resources on the fly. Another key benefit is the immutable event log they create, which allows for replayability. If something goes wrong, or if you need to audit decisions, you can rewind the stream to test new logic or recover from failures, which builds a lot of resilience into the system.
In the context of security, how can streaming agents contribute to real-time anomaly detection?
Streaming agents are incredibly powerful for security because they can analyze high-velocity data streams—like system metrics, network traffic, or sensor data—in real time to spot unusual patterns. By integrating metadata from incident records or threat feeds, they can cluster related events and feed that into AI workflows to pinpoint root causes or route alerts to the right teams. This cuts down the time to resolve issues significantly. The most useful data streams for this are often those with high frequency and granularity, such as login attempts or transaction logs, because they reveal deviations from the norm almost instantly.
What is your forecast for the future of agentic AI in production environments over the next few years?
I’m optimistic about the trajectory of agentic AI, but I think we’re still in the early stages of mastering production deployments. Over the next few years, I expect to see a stronger focus on data quality and real-time integration as foundational elements—organizations that get their data estate in order will lead the pack. We’ll also see more sophisticated event-driven architectures becoming standard, making AI agents more responsive and reliable. Industries like finance and telecom, which are already comfortable with complex systems, will likely push the boundaries with innovative use cases, while others will catch up by prioritizing human-in-the-loop models to build trust. Ethical considerations will also take center stage, ensuring AI doesn’t just perform well but does so responsibly.