Featured Work & Research
Architecting scalable AI solutions across FinTech, enterprise NLP, and edge computing.
HybridCredit-LLM (Institutional Risk Engine)
Active Research & Development
The Pitch:
Multimodal credit risk platform fusing LightGBM and 4-bit QLoRA fine-tuned Mistral-7B / Llama-3.3-70B for Basel III compliant risk modeling.
Tech Stack:
Mistral-7B, LightGBM, QLoRA, SHAP, Scikit-learn, Python
Features:
SHAP TreeExplainer, ECOA Fair Lending demographic audits, custom Flask Glassmorphism UI with live LLM streaming.
Key Highlights:
Architecture: Dual-branch system combining a LightGBM classifier with a fine-tuned Mistral-7B-Instruct model (4-bit QLoRA) for minimal compute overhead.
Performance Metrics: 0.985 AUC-ROC (0.9845 AUC / 0.9693 PR-AUC) vs 0.6709 LightGBM Baseline.
Enterprise Compliance: Fully aligned with the EU AI Act (Art. 13-15), ECOA, Basel III IRB, and GDPR Art. 22 standards through demographic fairness audits and a multi-modal explainability pipeline (SHAP + Counterfactuals).
Enterprise Hybrid RAG Engine (SEC 10-K QA)
Active Research & Development
The Pitch:
Multi-stage retrieval: Dense (ChromaDB) + Sparse (BM25) with Reciprocal Rank Fusion & Cross-Encoder reranking for Llama 3 QA.
Tech Stack:
Mistral-7B, LightGBM, QLoRA, SHAP, Scikit-learn, Python
Features:
Automated inline citation verification preventing financial report hallucinations.
Key Highlights:
Architecture: Query → [Dense Search + Sparse Search] → RRF Fusion → Cross-Encoder Rerank → LLM Generation.
Performance Metrics: RF Fusion | Cross-Encoder Reranking | Citation Validation | LLM-as-Judge Scoring.
Enterprise Compliance: Enterprise-grade Hybrid RAG Engine for SEC 10-K financial QA. Features Dense + BM25 retrieval, Reciprocal Rank Fusion (RRF), Cross-Encoder reranking, grounded Llama 3 generation, citation validation, and LLM-as-a-Judge evaluation.
Enterprise NLP Engine: Smart Email Classification System
Completed
The Pitch:
An intelligent text-processing pipeline designed to automate corporate communication sorting. Trained on a massive enterprise dataset, this system accurately routes high-volume communications into actionable categories like Business, HR, Legal, and Finance, drastically reducing manual administrative overhead.
Tech Stack:
Bidirectional LSTM, TensorFlow/Keras, Streamlit, Gmail API, Pandas
Key Highlights:
Deep Learning Pipeline: Built a Bidirectional LSTM neural network trained on the Enron Email Dataset containing over 500,000 records.
End-to-End NLP: Developed a comprehensive architecture encompassing automated text cleaning, tokenization, padding, and advanced word embeddings.
Interactive Deployment: Integrated the live Gmail API to dynamically fetch real-time data, visualizing label distributions and confidence scores via a custom Streamlit dashboard.
Intelligent Edge Automation System (SSIP Funded Rs. 1.2 Lakh - Gesture Pilot)
SSIP-Funded Prototype (Market-Ready)
The Pitch:
A real-time computer vision and automation architecture deployed directly onto edge hardware. This project successfully bridged the gap between complex software algorithms and physical hardware constraints, securing government backing to transition from an academic prototype to a commercially viable product.
Tech Stack:
NVIDIA Jetson Nano / Orin NX, CUDA, OpenCV, RTSP, C++ / Python
Key Highlights:
Government Backing: Awarded ₹1,20,000 in competitive funding through the Student Startup and Innovation Policy (SSIP) and the Government of Gujarat.
Hardware Integration: Led the end-to-end architecture and deployment on NVIDIA Jetson hardware.
System Optimization: Utilized CUDA for real-time edge processing and integrated RTSP streaming, increasing overall cost efficiency by 25% and preparing the underlying architecture for a pending IPR patent.