Pranjal Sinha

Conversational Multi-Agent BI System: Architected a LangGraph orchestrator coordinating dual deep-research agents, internal (hybrid dense+BM25 retrieval, SQL agent, session memory) and external (domain-whitelisted search, claim extraction with cross-reference validation) feeding a Tree-of-Thought synthesis engine with beam search, 4-dimension evaluation scoring, and an RLHF feedback loop that auto-tuned weights from edit patterns and downstream outcomes. Cut report TAT 94%. LLM Foundation Model: Domain-adapted a 7B LLM for cross-market credit risk scoring via LoRA fine-tuning; achieved 0.75 AUC with ~100 target-market samples, cutting new-market deployment time by 85%. Served at p95 ~120ms using 4-bit quantization + vLLM, with calibration layer for reliable probability outputs. Neuro-Symbolic Fraud Detection: Architected a hybrid GNN and RuleFit system for real-time fraud prevention. Achieved a 73% increase in detection rates (Spain). High-Performance RAG pipeline: Engineered a retrieval pipeline using FinBERT and vector search to perform transaction mining on unstructured bank statements. Optimized the indexing pipeline for low-latency retrieval, directly improving the accuracy of downstream credit scorecards by 23%. Contributed the FIGU (Fast Interpretable Greedy Unions) algorithm to the imodels open-source library.

Geospatial Deep Learning (Satellite Imagery): Developed a proprietary pipeline using CNNs and Time Series Forests to estimate farmland seasonality and yield. Enabled underwriting in data-scarce regions, increasing farming loan approval rates by 23%. Orthogonal Neural Network for Risk: Designed a Neural Network utilizing "lifestyle data" features orthogonal to standard bureau data. Solved the "cold-start" problem for New-To-Credit (NTC) customers, driving a 33% lift in credit card approvals. Bank Statement Parsing Engine: Engineered a raw-text parsing engine utilizing NLP to extract structured financial features from Account Aggregator data. Deployed as a microservice (logistic regression), improving Personal Loan approval rate by 17%. ML Observability: Built a comprehensive MLOps monitoring suite (drift detection, performance benchmarking) for 3 live production models, ensuring data integrity and minimizing model degradation downtime.

Analytics Infrastructure & Growth: Architected a data analytics pipeline to track interaction metrics across 500,000+ users. Engineered automated dashboards that drove content strategy, resulting in 40% YoY growth and a 35% increase in organic reach. Monetization & Scale: Leveraged user behaviour data to optimize affiliate conversion funnels (OYO), securing 1.1M+ in revenue. Scaled platform reach to 3,000+ colleges (IITs/IIMs) through data-backed partnership strategies.