Gen AI Engineer
COGNIZANT · Jul 2023 – Present
- Built a GenAl-powered document intelligence pipeline to process personal loan applications, automating extraction and validation of bank statements, pay stubs, government-issued ID, Social Security verification, and employment proof.
- Designed an end-to-end pipeline covering document ingestion, OCR/Document Al extraction, LLM-based structuring, data validation, and handoff to loan officers.
- Enabled the GenAl layer to identify salary, employer, monthly obligations, existing loan payments, bank balances, bounced payments, inconsistencies, and missing documents.
- Generated concise, structured loan-application summaries for the credit team, reducing manual document-review effort and turnaround time.
- Built a Retrieval-Augmented Generation (RAG) copilot on Azure OpenAl (GPT models) and Claude models to assist credit analysts and underwriters processing 40-50 applications daily, supporting rather than replacing the final credit decision.
- Unified applicant data, credit bureau data, bank statements, existing loans, internal credit policy, and product rules into a single retrieval context, using a document chunking strategy tuned for policy and financial documents to improve retrieval precision.
- Used RAG to retrieve relevant underwriting policies and an LLM to generate structured risk assessments, including risk level, debt-to-income ratio checks, credit score policy checks, delinquency flags, and recommended next actions.
- Evaluated copilot output quality using RAGAS metrics (faithfulness, answer relevancy, context precision) alongside an LLM-as-a-Judge framework, continuously validating retrieval and generation quality before production release.
- Delivered explainable, policy-grounded risk summaries that accelerated underwriter decision-making while keeping a human in the loop.
- Designed a compliance-constrained, multi-agent early-stage collections system for accounts with missed loan payments across a portfolio of thousands of active loans and credit cards.
- Evaluated payment history, days past due, outstanding amount, customer communication history, prior promises-to-pay, and risk segment to determine the next permitted action.
- Automated a compliant, day-by-day collections cadence, including payment reminders, payment status checks, approved follow-up communications, and eligible payment-assistance offers, with automatic escalation of higher-risk accounts to a human collections officer.
- Built a multi-agent architecture consisting of a Collections Orchestrator, Delinquency Analysis Agent, Customer Segmentation Agent, Communication Agent, Payment/Promise-to-Pay Agent, and Escalation Agent, using LangGraph and CrewAl to define agent state, task-delegation, and escalation logic within compliance guardrails.
- Designed system-level and task-level prompts aligned with banking compliance standards, embedding guardrails to prevent hallucination, scope drift, and non-compliant or out-of-policy responses.
- Implemented structured JSON-output enforcement and input/output validation for underwriting and risk workflows to ensure downstream systems received consistent, machine-readable results.
- Implemented LangFuse and LangSmith for end-to-end LLM observability, tracking prompt performance, response quality, latency, and token usage to surface hallucinations and optimize reliability in production.
- Established an LLM-as-a-Judge evaluation workflow paired with the RAGAS library to continuously score prompt and RAG changes before rollout, reducing regressions in production.
- Conducted iterative prompt experimentation and response scoring to optimize accuracy, reduce operational risk, and continuously tune guardrail effectiveness.
- Adopted Al-assisted coding tools, including GitHub Copilot and Claude Code, to accelerate feature development, test-case generation, and code review across GenAl services.

