03 · Projects
Production AI Systems.
Featured Project
CollectionIQ
Self-serve portfolio intelligence platform built for collection leaders. Plain-English query engine over LCC data so leaders get answers in under 30 seconds instead of waiting hours for analyst reports. Two LangGraph pipelines: a 4-agent query pipeline and an automated report pipeline with HTML generation and SMTP email dispatch. Used daily by 10+ collection leaders, zonal heads, and regional heads at Shriram Finance, supporting portfolio monitoring across 147+ branches and 14 regions.
⚡ 3-hour analyst workflow reduced to a 30-second self-serve query
Landing Page
KPI Dashboard
Smart Risk Alerts
AI Query Assistant
AutoCAM
Automates CIBIL report extraction from CRIF and TransUnion PDFs for credit analysts. Rule-based regex parsing with self-validation, LangChain and Gemini as fallback for edge cases. Outputs formatted Excel with DPD colour coding, eliminating manual data entry from credit appraisal workflows.
⚡ Analyst processing time reduced from 60 minutes to under 1 minute per borrower case
LCC Consolidator
Consolidates RE-wise LCC PDFs into a Master Excel for branch managers across Shriram Finance. Eliminates the manual monthly copy-paste process that previously took hours. Deployed live and actively used by branch teams.
⚡ Monthly consolidation effort reduced by 97% across all regions
Vehicle Insurance Prediction
End-to-end production grade MLOps pipeline predicting whether a customer opts for vehicle insurance. Full ML lifecycle from data ingestion to cloud deployment: MongoDB Atlas for data storage, AWS S3 as model registry, ECR for container storage, EC2 for hosting. CI/CD via GitHub Actions with a self-hosted EC2 runner. FastAPI interface for real-time predictions and training triggers.
⚡ Full MLOps lifecycle · automated CI/CD · Docker on AWS EC2
Blog Writing Agent
Multi-agent pipeline that mirrors how an experienced technical writer works. A Router classifies the topic and decides whether web research is needed. A Research node gathers Tavily results, deduplicates by URL, and produces a clean evidence pack. An Orchestrator plans each section with Pydantic-enforced contracts. Workers then write all sections in parallel via LangGraph Send, each grounded to verified citations only. A Merge node restores section order and assembles the final Markdown document.
⚡ Agentic routing · fan-out parallelism · schema-enforced section contracts