Vikas Mishra, a platform cloud architect at Google Cloud, shared his insights on the complexities of deploying AI brokers in manufacturing at MLDS 2025, India’s greatest builders summit hosted by AIM Media Home
Mishra mentioned that whereas the potential of enormous language fashions (LLMs) is broadly recognised, translating that into real-world functions presents vital hurdles. “2025 goes to be the 12 months of manufacturing,” he predicted, highlighting the growing strain to maneuver past theoretical functions and implement tangible options.
“You might have a number of brokers for each your inside and exterior use instances—issues like your buyer agent, worker agent, CX agent, and your code agent,” he mentioned.
Mishra outlined the important thing challenges that hinder profitable LLM deployment: efficiency, value, latency, and security. He mentioned that “solely about 50% of persons are truly operating them in manufacturing,” regardless of a a lot greater intent.
He argued that this hole stems from difficulties in evaluating, debugging, and guaranteeing the continual performance of brokers, particularly inside advanced multi-agent workflows. A minor change in immediate or mannequin behaviour can negatively influence person expertise, creating a necessity for sturdy analysis and monitoring instruments.
Google Cloud Vertex AI to the Rescue
Mishra confused the significance of choosing the suitable platform. “It’s not concerning the fashions,” he declared, “it’s concerning the platform.” He added that selecting the correct platform solves about 70% of the client’s issues.
He launched Google Cloud’s Vertex AI as a unified platform that addresses the challenges of productionising AI brokers. Vertex AI gives entry to a variety of fashions, together with Gemini, and gives instruments for tuning, internet hosting, and managing them.
Mishra mentioned that Google Cloud’s Mannequin Backyard, a platform for constructing, testing, and deploying fashions, hosts over 160 LLMs, together with each closed- and open-source fashions.“Fashions like Llama, Claude, and DeepSeek are all obtainable on Mannequin Backyard,” he mentioned.
Talking of Google’s Gemini and its distinctive capabilities, together with a “2 million context window” that permits the processing of huge quantities of knowledge, he showcased a reside demonstration of the mannequin’s multimodal capabilities, interacting with it by way of each audio and visible inputs.
RAG in Brokers
Mishra mentioned the essential position of grounding LLMs with real-time data. “You need entry to real-time information,” he mentioned, noting that Vertex AI permits grounding with Google Search and third-party information sources.
He additionally touched upon the method of mannequin adaptation, from immediate design to full coaching, and launched varied agent improvement instruments, together with Agent Builder, which gives low-code and no-code choices.
Mishra emphasised that agent analysis is essential however tough.
He described Vertex AI’s on-line analysis service and launched ‘Autorator’, a choose mannequin used for AI benchmarking. He additionally addressed the significance of mannequin observability and tracing, highlighting using Cloud Hint and Cloud Logging for debugging multi-agent workflows.
Lastly, he addressed the essential points of safety and governance. “You personal your information,” he mentioned, assuring the attendees that Google Cloud doesn’t use buyer information to coach its fashions. He additionally emphasised the significance of security filters and content material moderation instruments.
He concluded by reiterating that Vertex AI gives a complete platform for taking AI brokers from prototype to manufacturing, leveraging Google Cloud’s infrastructure and experience.
The put up Google Cloud’s Vikas Mishra: ‘2025 Is the Yr of Manufacturing’ for AI Brokers appeared first on Analytics India Journal.
