Featured Projects
AI Study Video Generator
Problem: Creating high-quality educational videos normally requires scripting, narration, animation, and video editing to be handled separately, making the process time-consuming and difficult to scale.
Solution: Built an agentic AI system that transforms a learning topic into a complete narrated explainer video through a multi-stage AI pipeline combining LLM planning, local text-to-speech, programmatic animation, and automated video assembly.
- 🤖 Four specialized agents orchestrated through LangGraph for planning, narration, visualization, and assembly.
- 🧠 LLM-based lesson planner creates a 3–6 shot pedagogical flow with a consistent example dataset.
- 🔊 Piper TTS generates fully local narration and measures the actual audio duration for animation synchronization.
- 🎨 Visual Agent generates Manim animation code and automatically retries failed renders using the rendering error.
- ⏱️ Audio-aware animation keeps generated visuals aligned with the actual narration duration.
- 🎬 FFmpeg automatically combines individual shots, narration, and visuals into the final educational video.
- ⚙️ Asynchronous job processing with SQLite-based status tracking for long-running video generation.
AI Study Companion
Problem: Students studying complex material often lack a way to get grounded, trustworthy answers and identify their own weak spots as they learn.
Solution: An agentic AI study assistant — built on a LangGraph ReAct agent with Retrieval-Augmented Generation (RAG) — that answers questions grounded in course material, generates quizzes, tracks weak topics, and evaluates its own answer faithfulness.
- LangGraph-based ReAct agent architecture for multi-step reasoning and tool use
- Retrieval-Augmented Generation (RAG) grounds answers in real course material, reducing hallucination
- Automatic quiz generation from studied content
- Weak-topic tracking to surface knowledge gaps over time
- LLM-as-judge faithfulness evaluator that scores its own answers for groundedness
MetroVerse AI
Problem: Metro commuters rely on multiple apps for route planning, ticketing, crowd information and station guidance.
Solution: Built a unified AI-powered metro assistant integrating smart routing, ML crowd prediction, conversational AI, live navigation and digital ticketing into one platform.
- Dijkstra-based shortest path routing across 262 stations on 11 metro lines
- ML-based crowd prediction model integrated across live routes
- Groq Llama 3.3 powered AI assistant with session memory
- QR ticketing, metro card management and real-time journey navigation
- Push notifications, QR station scanner and nearby tourist guide
VibeVerse
Problem: Students face anxiety and loneliness in new environments without personalized emotional support.
Solution: An AI-powered companion that understands sentiment and provides tailored wellness tools.
- Real-time mood detection using TextBlob sentiment analysis
- Mood-based personalized music recommendations via the Spotify API
- Daily journaling and streak tracking to build confidence
Smart College Assistant
Problem: Managing college resources and getting academic assistance was scattered and inefficient for students.
Solution: A centralized smart web platform to streamline student interaction with academic resources.
- Bilingual Hindi/English NLP chatbot trained on 500+ QA pairs using TF-IDF and a Voting Classifier (SVM + Logistic Regression + Random Forest)
- Responsive, student-focused interface for seamless navigation
- 10+ campus modules (chatbot, attendance, marks, dashboard) unified into one Flask REST API backend
Fake Certificate Detection ML
Problem: Manual certificate verification is slow and error-prone, letting forged documents slip through undetected.
Solution: A Machine Learning model that automatically classifies and detects manipulated certificates based on data patterns.
- Trained a Random Forest classifier on 34,000+ certificate records, achieving 84% classification accuracy
- Automates candidate evaluation & document screening pipelines
- Rigorous data preprocessing & feature engineering to boost model accuracy
House Price Prediction
Context: One of my first ML projects — built to learn how regression models actually work on real data.
What I did: Started with a simple Linear Regression baseline, then tried Decision Tree and Random Forest models to see how they compared.
- Learned to evaluate models using MAE and R² score instead of just guessing
- Saw firsthand how a Decision Tree overfits and how Random Forest fixes that
- Found that bathrooms, bedrooms, and sqft_living mattered most for predicting price