Vivek Singh
Cisco System, Sr Technical Leader Customer Experience
Pune, India
Actions
I am Sr Technical Leader in Cisco Customer Experience for Collaboration Technology.
Currently focused on Gen AI-based innovation and its impact on the customer experience within Cisco Org.
I have been a speaker in Cisco Live and have about 2 patent issued related to AI
Links
Badges
Area of Expertise
Topics
The Inquisitive Agent: Curing RAG's Ambiguity with Autonomous Dialogue
For all their power, most RAG systems today are locked in a monologue. They listen to a query, retrieve what they think is relevant, and speak their answer. If the initial query is vague, the monologue fails. The future of intelligent assistance, however, is not a monologue; it's a rich, collaborative dialogue.
This session dives deep into the architecture of AID (Assisted Intelligent Dialogue), an Agentic system we built and deployed at Cisco to make this future a reality. We've designed an system that doesn't just search for answers, but actively seeks clarity. Instead of a linear pipeline, our agents operates in a cognitive loop, embodying different roles to deconstruct ambiguity and co-create understanding with the user.
We will walk you through its autonomous reasoning process:
1. The Agent as the Detective (Ambiguity Assessment):
When a query arrives, the agent’s first task is to assess the evidence. It doesn't immediately jump to conclusions. Instead, it analyzes the query for semantic ambiguity and potential for misinterpretation, identifying cases where a simple vector search would likely fail.
2. The Agent as the Strategist (Hypothesis and Reasoning):
Upon detecting ambiguity, the agent consults its "knowledge model"—a rich knowledge graph. Here, it maps the user's query onto the graph to explore interconnected concepts and relationships. It generates multiple hypotheses about the user's true intent and identifies the critical piece of missing information needed to validate one of them. This is the agent’s core reasoning phase.
3. The Agent as the Interviewer (Probing with Purpose):
Armed with this insight, the agent reaches its "moment of truth." It plans and executes a dialogue action, formulating a targeted, clarifying question. This isn't a generic "Can you rephrase?"; it's a precise probe designed to efficiently collapse the space of uncertainty and guide the user toward the correct context.
4. The Agent as the Expert (High-Fidelity Resolution):
Once the user responds, the agent integrates this new, crucial context. With the ambiguity resolved, it can now perform a high-fidelity retrieval from a precisely defined subgraph, delivering an accurate, actionable answer that solves the user's true underlying problem.
Join us to learn how to build agents that don't just answer, but actively listen, reason, and understand. This is the blueprint for the next generation of truly helpful, conversational AI.
What Attendees Will Learn:
a)Design Pattern: How to architect an Agentic "clarification loop" to move beyond brittle, single-shot RAG systems.
b)Cognitive Architecture: How to leverage a graph as a "world model" for an agent to reason about information and generate hypotheses.
c)Practical Implementation: Techniques for an agent to autonomously decide when to ask a question and how to formulate it with purpose.
d)Human-Agent Interaction: Principles for designing effective, low-friction dialogues that guide users without causing fatigue.
e)From Prototype to Production: Lessons learned from deploying this Agentic system at enterprise scale at Cisco to solve daily customer issues, complete with a look at our evaluation metrics.
Agents as Data Engineers: Using LLMs to Query and Synthesize Scattered Service Data
For any enterprise service organization, the answers to critical questions are buried in a chaotic archipelago of disparate data sources: CRM databases, ticketing systems, application logs, and more. The real challenge isn't analysis; it's the painstaking, manual process of fetching, joining, and making sense of this scattered data.
This talk demystifies this process by introducing a two-stage, multi-agentic workflow that automates data engineering and interpretation.
Stage 1: The Data Forage. We'll show how an Orchestrator Agent breaks down a high-level business question (e.g., "Analyze ticket trends for our top clients") into sub-tasks.
Stage 2: The Sense-Making Layer. Raw data is just noise. The second, and most crucial, stage involves agents designed to understand the business. We'll explore how a Business Context Agent takes the raw, structured data from the forage and applies organizational knowledge. It resolves ambiguities, understands what defines a "top client" based on database fields, maps customer_id from one system to org_id in another, and ultimately transforms the raw data into a coherent, analysis-ready dataset.
The GraphRAG System That Asks Back: Intelligent Probing with Better Contextual Answers
In customer support and technical assistance, resolving user issues is often hindered by incomplete or ambiguous queries. Traditional RAG systems, relying heavily on semantic similarity, struggle to fully grasp the user's intent from vague input, leading to suboptimal answers. This session introduces AID (Assisted Intelligent Dialogue), a sophisticated knowledge system developed at Cisco specifically to overcome these limitations and tackle real-world customer challenges.
We will explain how AID enhances standard RAG by leveraging Neo4j graphs to build a deeper multimodal understanding of content and its relationships. A key innovation is AID's intelligent probing mechanism. By using techniques, including generating hypothetical questions, AID first refines the search space within the knowledge graph to identify the most relevant contexts. From these precise subgraphs, AID formulates clarifying questions, effectively "asking back" the user to gather crucial missing context and refine their intent through an assisted intelligent dialogue.
You will learn how this graph-powered approach directly improves Cisco's ability to solve day-to-day customer issues by transforming unclear inquiries into well-defined problems, resulting in significantly more accurate and actionable contextual answers. The session will cover AID's architecture, its multimodal retrieval, dynamic metadata extraction, and, critically, how Neo4j enables this intelligent probing and dialogue to deliver a superior user experience and enhance support outcomes. Join us to discover how to build AI systems that intelligently interact to solve complex problems, ensuring your users get the precise information they need.
NODES 2025 Sessionize Event
AgentCon Pune Sessionize Event
AgentCon 2025 - Hyderabad, India Sessionize Event
AgentCon 2025 - Washington Sessionize Event
Vivek Singh
Cisco System, Sr Technical Leader Customer Experience
Pune, India
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top