The latest articles on DEV Community by sanjjanakoyalkar-cmd (@sanjjanakoyalkarcmd). https://dev.to/sanjjanakoyalkarcmd https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4147545%2F0f1b17db-9722-4e6f-b2fa-71d829d97435.png https://dev.to/sanjjanakoyalkarcmd en sanjjanakoyalkar-cmd Mon, 28 Sep 2026 16:03:20 +0000 https://dev.to/sanjjanakoyalkarcmd/building-an-escalation-memory-agent-because-angry-customers-shouldnt-wait-47ea https://dev.to/sanjjanakoyalkarcmd/building-an-escalation-memory-agent-because-angry-customers-shouldnt-wait-47ea <p>When a customer sends an angry email or threatens to churn, support agents scramble. They search old tickets, check billing, look at usage, and try to remember what worked last time. Meanwhile, the customer gets angrier.</p> <p>For the HackwithHyderabad hackathon, we built an <strong>Escalation Memory Agent</strong>: it remembers each customer's journey and briefs the support agent instantly.</p> <h2> What the agent remembers </h2> <ul> <li>Tickets and resolutions from the last 6 months</li> <li>Product usage trends (login frequency, feature adoption)</li> <li>Billing history</li> <li>Past escalations and which recovery tactic worked</li> <li>The customer's communication style and preferred channel</li> </ul> <h2> What the agent gives you </h2> <p>A brief with a risk score (HIGH / MEDIUM / LOW), the reasons behind it, and a recommended action. Every claim links to its source, like "Ticket T-1155", so agents can verify it.</p> <h2> Design choices </h2> <ol> <li> <strong>Facts come from data, not from a language model.</strong> The risk engine is deterministic, so it can't hallucinate a ticket that doesn't exist.</li> <li> <strong>Learning loop.</strong> After an escalation, the agent logs whether the tactic recovered the customer. Next time, it recommends tactics based on real success rates, first for that customer, then for customers with a similar communication style.</li> <li> <strong>Persistent memory.</strong> Add a new ticket and the next brief changes.</li> </ol> <h2> What's next </h2> <p>An LLM layer to draft the reply, retrieval over ticket text, and live Zendesk and Stripe connectors.</p> <p>Demo video: <a href="https://youtu.be/lLKUMVi1Z60" rel="noopener noreferrer">https://youtu.be/lLKUMVi1Z60</a><br> Try it: <a href="https://sanjjanakoyalkar-cmd.github.io/escalation_memory_agent/" rel="noopener noreferrer">https://sanjjanakoyalkar-cmd.github.io/escalation_memory_agent/</a><br> Code: <a href="https://github.com/sanjjanakoyalkar-cmd/escalation_memory_agent" rel="noopener noreferrer">https://github.com/sanjjanakoyalkar-cmd/escalation_memory_agent</a></p> ai hackathon agents saas