We are building an AI-native analytics function. Instead of analysts hand-writing every query and report, we want people who design AI agents that explore lending data, find insights and recommend actions on their own. You will connect LLMs to our data warehouse and internal tools. You will write the prompts, guardrails and evaluation loops that make the agents reliable.
You don't need to be a lending expert. You need working knowledge of how lending and collections work, and deep skill in building and prompting agents. We'll teach the domain. We can't teach you to build valuable agents from scratch.
What You'll Do
• Design, build and ship LLM-based agents that query Snowflake, analyse portfolios and produce insights, such as roll rates, bucket movement, cohort performance and campaign effectiveness.
• Write and maintain production-grade system prompts, tool definitions and knowledge bases, so agents apply business rules correctly (for example, block-list filters, DPD definitions and loan-prefix logic).
• Build multi-step and multi-agent workflows: planning, SQL generation, validation,
summarization and anomaly flagging.
• Create evaluation harnesses (golden questions, SQL-correctness checks, hallucination tests) to measure and improve agent accuracy.
• Automate recurring analyses: daily and weekly performance reports, pre-due and delinquent campaign tracking, field-visit and agency productivity analysis.
• Turn agent output into dashboards, artifacts or alerts that collections, risk and product teams use.
• Work with business stakeholders to turn vague questions ("why did the 30+ bucket worsen?") into agent-solvable tasks.
• Document agent behaviour, failure modes and costs, and keep token and compute usage under control.
Must-Have Skills
• Agent building: hands-on experience with tool-use / function calling and agent loops, using the Claude API/SDK, MCP, LangChain/LangGraph or similar.
• Advanced prompting: structured prompts, few-shot examples, chain-of-thought, o
📌 AI Agent Engineer (Mumbai)
🏢 Kissht
📍 Mumbai