2026
Interactive Hardware-Aware · Deep learning
NeuraLens
ML and hardware engineers have no single resource that bridges mathematical intuition, live computation, and silicon-level hardware analysis for deep learning layers. Textbooks give static equations. Playgrounds have no hardware context. FPGA guides have no ML intuition. Everything is scattered.
An interactive, single-file educational web application for exploring, visualizing, and understanding deep learning layers with real-time hardware analysis, live visualizations, and multi-level learning content designed for ML engineers, AI researchers, and hardware architects.
HTML5 Canvas APIVanilla JavaScript ES2020+
2026
Machine Learning · Customer Management
ShieldGraph
E-commerce platforms lose millions in logistics costs to organized Return-to-Origin (RTO) fraud rings placing malicious, anonymous Cash-on-Delivery orders with zero intent to pay.
I built a hybrid GNN and Transformer model that uncovers coordinated fraud rings via shared technical infrastructure and behavioral click-streams in under 10 milliseconds. The system intercepts live checkouts to dynamically disable Cash-on-Delivery for risky accounts, forcing digital prepayment to break the fraud ring's financial incentives.
Real-Time APIsGraph AnalyticsPyTorch Geometric (PyG)Graph Neural Networks (GNN)
2026
Machine Learning · Credit Risk
CreditRisk AI
Banks lose money when 21.8% of borrowers default on loans. Traditional credit scoring fails for MSMEs and individuals with no credit history, and manual underwriting is slow and inconsistent.
Built an XGBoost model that analyzes 10 features from loan applications (income, loan amount, interest rate, credit grade, employment history) to predict probability of default. The model approves safe borrowers (PD < 30%), reviews moderate risk (30-50%), and rejects high risk (> 50%), while providing top 3 reasons for each decision using SHAP values for regulatory compliance. Deployed as FastAPI REST endpoint with Docker containerization.
XGBoostSHAPPython
2026
Machine Learning . AgriTech. Open Source
Groundwater Depletion Risk Prediction in India
India's agriculture relies heavily on groundwater, yet increasing extraction, irregular rainfall, and growing water demand are causing aquifers in many regions to deplete faster than they can naturally recharge. This threatens crop productivity, farmer incomes, rural water security, and the long-term sustainability of agricultural communities.
This project successfully developed a machine learning model to predict groundwater
depletion risk in Indian agricultural regions.Haryana village demonstration, where extraction (7.5 BCM) exceeded the extractable
resource (6.0 BCM) and the model correctly predicted High Risk, validated the system’s
utility as a budgeting and early-warning tool for smallholder farmers.
0.93–0.98 F1 Scores
Across all 3 classes
NumPyXGBoostPythonSeabornscikit-learn
2026
Agentic AI · Data Analytics
Product Performance Retrieval Agent
Raw financial CSV datasets are opaque to non-technical stakeholders — critical insights buried in numbers with no pathway to automated reasoning or visualization.
Built a modular LLM-driven pipeline: a dataset profiler extracts metadata (types, cardinality, distributions), feeds a Groq LLaMA 3.1 reasoning layer that dynamically suggests 6 tailored insights, then routes them to deterministic Python analysis functions. Results are auto-visualized via Streamlit, scored for data quality, and made queryable through interactive follow-up Q&A.
LLaMA 3.1Groq APIStreamlitPandasRAG
2026
RAG · Data Analytics
Legal Document Intelligence System
Raw financial CSV datasets are opaque to non-technical stakeholders — critical insights buried in numbers with no pathway to automated reasoning or visualization.
Developed RAG-based pipelines to process large legal PDFs with structured retrieval workflows. Implemented chunking, embedding, and similarity search to improve response accuracy and optimized retrieval
logic to support multi-document reasoning tasks.
LLaMA 3.1Groq APILangChaiPandasRAG
2026
Reasoning AI · Graph
Semantic Diffing for Evolving Knowledge Graphs
Raw financial CSV datasets are opaque to non-technical stakeholders — critical insights buried in numbers with no pathway to automated reasoning or visualization.
Built LLM-powered knowledge extraction pipelines generating temporal graphs from legal datasets. Engineered semantic diffing logic to detect contradictions across evolving documents and enabled fast
entity-relation reasoning across large datasets.
LLaMA 3.1Groq APIStreamlitPandasRAG
2024
Computer Vision · Safety AI
Intelligent Brawl Monitor
Violence escalates without early warning, limiting timely intervention across public safety, workplace security, and women's safety scenarios.
Real-time video analytics pipeline using OpenCV for frame capture. CNN-based emotion models detect anger-related expressions; pose estimation (MediaPipe/OpenPose) identifies aggressive body stances. Rule-based fusion logic fires alerts only when both signals align — minimizing false positives. Automated email notification system for escalation.
OpenCVMediaPipeCNNTensorFlowEmail Alerts
2025
Computer Vision · Re-ID
Cross-Camera Human Identification System
Multi-camera human tracking in dynamic environments fails to maintain consistent identities across views without facial recognition — limiting surveillance scalability and privacy compliance.
View-invariant pipeline: YOLOv8 detects humans, BoT-SORT assigns per-camera track IDs, color histogram embeddings + temporal trajectory encoding enable cross-camera matching via Bhattacharyya distance & cosine similarity. Global ID map output as annotated MP4 per camera, original footage preserved.
YOLOv8BoT-SORTOpenCVCosine SimilarityNumPy
2025
NLP · Unsupervised ML
Cognitive Stress Analysis from Speech
Cognitive stress shapes speech patterns but is difficult to measure objectively without intrusive methods or labeled data.
End-to-end Python pipeline performing speech recognition, acoustic signal processing, and linguistic feature engineering to capture speech rate, pitch, pause patterns, hesitation markers, and sentence complexity. K-Means clustering + Isolation Forest anomaly detection + cosine similarity quantify and compare stress indicators. Auto-generates interpretable CSV reports.
librosaNLTKscikit-learnIsolation ForestK-Means
2025
Computer Vision · Research
Open-Vocabulary Object Detection Study
Traditional detection models are bound to predefined categories — evaluating zero-shot generalization and real-time speed vs. accuracy trade-offs across architectures remains an open challenge.
Side-by-side benchmark of OWL-ViT (zero-shot, text-query driven) vs. YOLOv8 on video feeds. Per-frame bounding box annotation, CSV-logged confidence scores, GPU-accelerated with mixed precision. Statistical analysis (T-tests, distribution plots) surfaces false-positive patterns and model divergence.
OWL-ViTYOLOv8HuggingFacePyTorchSeaborn
2025
Full-Stack AI · Healthcare
MediHelper
Reliable preliminary skin disease identification and healthcare cost estimation are hard to access — leading to delayed diagnosis and poor financial planning for patients.
Full-stack Django web application integrating two ML models: a CNN trained on skin condition datasets (vitiligo, acne, SJS, hyperpigmentation, nail psoriasis) for image-based disease detection, and a linear regression model for healthcare cost prediction from user-provided health parameters. Modular MVC architecture designed for extension (chatbot, medicine scanning).
DjangoTensorFlowCNNLinear RegressionPandas
.jpeg)
.jpeg)
.jpeg)
.jpeg)
