Project Details

● Developed an AI sneaker assistant for e-commerce, combining conversational search, FAQ handling, & product discovery in one interface.
● Scraped 2,000+ Myntra shoe listings using Selenium & BeautifulSoup, structuring prices, sizes, colors, ratings, discounts, & availability into a searchable catalog.
● Covered men’s & women’s footwear across 7+ brands, improving catalog breadth & recommendation relevance.
Implemented semantic search with Sentence Transformers & ChromaDB, enabling sub-second retrieval for natural language product queries.
● Engineered Semantic Router-based intent classification to route FAQ, product search, & small-talk requests to the right workflow.
● Translated natural language shopping requests into structured product lookups, returning user-friendly responses with product context.
● Integrated SQLite-based retrieval with Groq Cloud API & Ollama inference, supporting flexible cloud & local deployment.

Project video

Project Details

● Developed a fine-grained emotion recognition framework that predicts emotion probability distributions instead of forcing a single hard label, making the system better aligned with subjective human perception.
● Leveraged transfer learning with ResNet-50 & KL Divergence loss on the Emotion6 dataset to learn soft labels from multi-annotator emotion ratings.
● Achieved 61.62% Top-1 accuracy, 79.55% Top-2 accuracy, 0.16 KL Divergence, & 0.60 average Pearson correlation, closely matching published benchmark performance.
● Improved disgust detection to 0.73 Pearson correlation with human judgment while also delivering strong results for joy at 0.71 & sadness at 0.65.
● Evaluated alternative architectures including EfficientNet-B0 & ConvNeXt-Tiny to compare feature extraction quality, parameter efficiency, & model behaviour.
● Operationalized the solution through a FastAPI backend & Streamlit interface, enabling probability-based emotion analysis in about 2 seconds per image.

Project video

Project Details

● Developed production-ready 2-agent AI system using Google’s A2A protocol to separate academic paper retrieval from LLM summarization, improving modularity, interoperability, & fault isolation.
● Configured ResearchAgent to query the arXiv API across research papers, collect metadata, & delegate summarization work to SummarizerAgent.
● Implemented Agent Card discovery so agents could advertise capabilities using standardized endpoints, removing hardcoded dependencies & enabling dynamic capability matching.
​● Modeled a 3-state task lifecycle, SUBMITTED, WORKING, & COMPLETED, with Pydantic Tasks & typed Artifacts for structured data exchange, state tracking, & safer workflow execution.
● Enabled async HTTPX polling with retry logic for concurrent inter-agent communication, allowing distributed processing without blocking retrieval or summarization.
● Delivered Streamlit UI & CLI interfaces, producing 2 structured outputs for researchers: paper summaries & follow-up questions.

Project video

Project Details

● MedTrace is a biomedical RAG system that retrieves peer-reviewed PubMed literature, chunks & embeds articles into FAISS, and answers research queries with source-linked evidence synthesis.
● Connected to PubMed via NCBI E-utilities API, supporting semantic search over 39M+ articles and configurable ingestion of 10-299 papers per topic.
● Added Groq-based stance classification to label studies as Supporting, Opposing, or Neutral, surfacing contradictions & summarizing conflicting findings without manual meta-analysis.
● Reduced literature synthesis time by ~60% and achieved 80% agreement with manual validation across 50 research articles spanning 2 clinical domains.
● Integrated LLM-as-judge metrics, Relevance, Groundedness, & Confidence, into a real-time Streamlit dashboard with direct PubMed source linking for auditability.
● Reached 85%+ retrieval precision, giving users transparent answer quality signals instead of standard RAG outputs that ignore evidence consistency.

Project video

Project Details

● Healthline RAG Assistant is a medical research chatbot that automates article ingestion, semantic search, & grounded Q&A over Healthline content only.
● Processes 1–10 validated Healthline URLs per session through URL validation, canonicalization, & deduplication to improve retrieval reliability.
● Used LangChain, ChromaDB, & Hugging Face embeddings to build a local vector store for fast semantic retrieval from indexed articles.
● Integrated Groq API with optional Ollama inference to generate source-grounded responses constrained to retrieved context.
● Tested 50+ queries across multiple articles, achieving 100% grounded responses with zero hallucinations in evaluation.
● Enforced a strict fallback that returns “information not found” when evidence is absent, prioritizing accuracy & user trust over unsupported answers.

Project video

Project Details

● CredVibe is a credit risk assessment system that predicts default probability, assigns CIBIL-like scores from 300-900, and classifies loan applicants into 4 risk tiers: Poor, Average, Good, and Excellent.
● Built a Logistic Regression + Optuna model using 14 features across personal, bureau, and loan data, achieving 95.31% recall on defaulters, Gini >85, and KS >40 to meet financial risk benchmarks.
● Improved model robustness with IV-based feature selection, VIF filtering, derived features such as LTI and Delinquency Ratio, and out-of-time validation for temporal stability.
● Deployed real-time scoring through FastAPI, Streamlit, and MLflow, enabling explainable predictions, business rule integration, and sub-millisecond inference.

Project video

Project Details

● FreshCheck AI is a deep learning-based fruit quality assessment system that performs Fresh vs Spoiled classification across 8 fruit types, supporting faster quality checks and earlier inventory action.
● Built with PyTorch, ResNet-50, and Optuna, it uses transfer learning with a frozen backbone and fine-tuned classifier to deliver fast, confidence-scored predictions in a lightweight deployment setup.
● Implemented fruit-aware and class-aware dataset splitting to prevent data leakage and ensure evaluation reflects real-world performance on unseen samples instead of memorized training patterns.
● Deployed an interactive Streamlit app for real-time image classification, reducing manual inspection effort and enabling quicker spoilage screening with a production-first focus on data integrity and inference efficiency.

Project video

Project Details

● DentDetect AI is a smart vehicle damage assessment system that classifies car images into 6 categories: Front Normal, Front Breakage, Front Crushed, Rear Normal, Rear Breakage, and Rear Crushed.
● Trained on 3,000+ labelled vehicle images, the system uses ResNet-50 transfer learning with data augmentation to improve robustness across varied car damage patterns and viewpoints.
● Achieved 80.70% validation accuracy, improving over a baseline pretrained ResNet-50 model at 77.00% and substantially outperforming EfficientNet-B0 at 65.04%.
● Applied Optuna-based hyperparameter tuning on fully connected layers, dropout, and learning rate settings to improve classification performance without redesigning the full architecture.
● Deployed through FastAPI and Streamlit, enabling real-time image-based predictions with confidence scores for faster insurance claim review and repair assessment workflows.

Project video

Project Details

● BevIntel is an ML-powered beverage pricing system built to predict optimal price ranges from customer demographics, consumption habits, product preferences, and brand familiarity signals.
● Trained LightGBM and XGBoost models on 40,000+ records, transforming messy, siloed data into a unified pricing workflow for real-time prediction.
● Engineered business-driven KPIs such as zone influence score, consumer frequency, brand awareness, and Brand-Switching Indicator to capture regional and loyalty-based pricing behavior.
● Integrated MLflow and DagsHub for experiment tracking, model versioning, and reproducible inference across iterative model development.
● Delivered an interactive Streamlit app that enables business teams to test pricing scenarios instantly without manual calculations.

Project video

Project Details

● InsureSight AI is a health insurance premium prediction system that estimates policy costs in real time using demographic, medical, lifestyle, and coverage inputs.
● Built with XGBoost and Scikit-learn, it uses 2 age-segmented models, for users aged ≤ 25 and >25, to capture life-stage risk differences more effectively than a single pricing model.
● Standardized underwriting logic through a normalized 0-1 medical risk score covering 4 major conditions, improving consistency across premium estimates and customer profiles.
● Supports instant INR quote generation through a Streamlit app with input validation, categorical encoding, and joblib-based model loading for fast, stable inference.
● Incorporates 3 insurance plan tiers, 4 regions, and structured health-risk features, enabling more tailored premium estimates while reducing manual pricing effort for common quote scenarios.

Project video

Project Details

● SpendSensei is a full-stack expense management system that helps users record, update, delete, search, and analyze personal expenses through a MySQL-backed web app.
● Built with FastAPI and Streamlit, it supports categorized expense tracking, paginated transaction views, and note-based search for faster access to historical records.
● Added analytics dashboards for category, month, and day-of-week spending breakdowns, helping users identify patterns, monitor habits, and spot potential savings.
● Implemented duplicate prevention, strict input validation, predefined category controls, backend logging, and secure .env-based credential storage to maintain clean and reliable data.

Project video