ML-Driven Monocular Depth Estimation for 3D Spatial Perception

What I did: Built a perception system that converts 2D images into 3D spatial understanding using MiDaS depth models, supporting real-time webcam, batch processing, and structured 3D exports.

How I did it: Leveraged PyTorch with MiDaS (DPT_Large/Hybrid) for GPU-optimized depth estimation, OpenCV for image processing, and Plotly for 3D point cloud visualization. Deployed via Streamlit and Gradio web interfaces.

Multi-Agent Warehouse Simulator with Scheduling, Pathfinding & Collision Avoidance

What I did: Built a multi-robot coordination simulator with pluggable scheduling and pathfinding algorithms, collision avoidance, deadlock detection, and real-time visualization for warehouse automation.

How I did it: Implemented Strategy pattern for runtime-swappable algorithms (FCFS, Priority, Greedy scheduling + A*, Dijkstra pathfinding), reservation-based collision avoidance, and wait-for graph deadlock detection. Visualized with Pygame.

Hybrid Semantic Code Search with Bi-Encoder Retrieval & Cross-Encoder Reranking

What I did: Built a semantic code search engine combining dense (FAISS) and sparse (BM25) retrieval with cross-encoder reranking, supporting natural-language queries across multiple programming languages.

How I did it: Used Sentence-Transformers for bi-encoding, FAISS for vector indexing, BM25 for keyword retrieval, and RRF for result fusion. Built FastAPI endpoints with AST-based code chunking and SQLite metadata storage.

Time-Series Health Metrics API with Simulated Wearable Sensor Data

What I did: Built a backend service that processes 186K+ simulated wearable sensor events into computed health scores (Sleep, Strain, Recovery) via versioned REST APIs.

How I did it: Used FastAPI with async PostgreSQL (SQLAlchemy + Alembic), Docker Compose for deployment, and designed a sensor event pipeline with algorithm versioning and metric caching.

AI-Powered Public Startup Intelligence Platform with Automated Data Pipelines

What I did: Built a startup intelligence platform with AI chat, idea marketplace, and automated news aggregation using ML-powered data pipelines across 5 interconnected services.

How I did it: Developed Next.js 16 frontend, FastAPI + PostgreSQL + Weaviate backend, Apache Airflow data pipelines with Pegasus summarization, and ML-based article filtering. Deployed with Docker and GCP.

Transformer-Based Italian-English Neural Machine Translation System

What I did: Built a sequence-to-sequence neural machine translation system for Italian-English translation using a custom Transformer architecture with multi-head attention.

How I did it: Implemented the full Transformer encoder-decoder pipeline in PyTorch with self-attention, cross-attention, positional encoding, and autoregressive decoding trained on parallel corpora.

Structured Information Extraction from Financial Texts using Deep Learning

What I did: Built a transformer-based deep learning pipeline to extract Subject-Predicate-Object triplets from unstructured financial news using token-level BIO tagging.

How I did it: Utilized BERT, BiLSTM, and RNN-LSTM models in PyTorch, with spaCy for tokenization and sequence labeling, achieving high extraction accuracy through attention-enhanced modeling.

AI-powered video transcription and subtitle generation.

What I did: Developed TranscriBelt, a web application for AI-powered video transcription and subtitle generation, with CRUD operations on transcripts.

How I did it: Integrated OpenAI Whisper for audio-to-text processing and OpenCV for frame-level video parsing. Built scalable backend using FastAPI, Celery, Redis, and AWS (Lambda, S3, DynamoDB). Frontend developed in Next.js and TypeScript with Clerk.js for secure access.

Personalized AI flashcard generator for adaptive and assistive learning experiences.

What I did: Developed an adaptive learning tool that leverages LLMs to generate flashcards based on user-provided input, streamlining human-in-the-loop content curation.

How I did it: Built in Next.js and Tailwind CSS with TypeScript; integrated LLM prompts for flashcard synthesis and Clerk.js for secure user sessions. Stripe added for access-tiered services.

Smart pantry assistant integrating AI for object detection and recipe recommendation.

What I did: Built a mobile-first AI assistant that automates pantry tracking and proposes context-aware recipes based on user inventory and image inputs.

How I did it: Used GPT-4o for item detection from photos and LLaMA 3 for recipe generation. Developed frontend in Next.js and Material UI. Integrated Google OAuth for seamless sign-in.

Real-time facial image enhancement using generative priors for restoration tasks.

What I did: Developed a deep learning pipeline using GFPGAN to restore low-resolution facial images, targeting deployment in accessibility and photo-repair applications.

How I did it: Implemented the GFPGAN model using Python and PyTorch, and built a user-friendly web application with Gradio for real-time image processing and upscaling. Achieved a 4x resolution enhancement and 92% accuracy in facial detail restoration, with an average processing time of under 10 seconds per image.

Web-deployable face-swapping tool using SOTA face recognition and synthesis models.

What I did: Engineered a facial identity transformation app leveraging ONNX-optimized InsightFace models, focusing on precision and real-time user interaction.

How I did it: Deployed Buffalo_l and inswapper_128.onnx on Hugging Face Spaces with a Gradio UI; hosted inference API on AWS EC2 for sub-100ms response time.

A Versatile and Accurate Search Engine

What I did: Developed a search engine equipped to handle a wide array of search functionalities, ensuring precise and relevant results through advanced algorithms and techniques.

How I did it: Implemented tokenization, normalization, boolean and ranked retrieval, positional indexing, and supported large datasets with Variable-Byte Encoding and On-disk Indexing. Utilized text processing techniques like stemming, lemmatization, and stopword removal, supporting real-time indexing and multi-threaded query processing.

An iPhone Text Messaging Application to Transform your Voice

What I did: Developed and co-created Mock’EM, an innovative iMessage extension app (which ranked #68 in the App Store under Entertainment ) enables users to transform their voice or text into captivating deep fake voices of their favorite character or celebrity and send hilarious and engaging messages to friends through iMessage.

How I did it: Employed SwiftUI to design the frontend and integrated Swift along with the FakeYou API for the backend, resulting in a seamless and user-friendly application that offers fantastic deepfake generation capabilities.

A Food Delivery Aggregator & Price Comparison Application

What I did: Despite not being able to access any API or public data, I created a comprehensive price comparison app for popular food delivery services, such as GrubHub, DoorDash, UberEats, and Postmates to help consumers save when ordering their favorite food items.

How I did it: I utilized React Native with tailwind-rn for the frontend and Express JS for the backend to deliver a seamless and user-friendly application. I applied reverse engineering techniques to clone and retrieve data from the targeted food delivery service APIs which do not exist publicly, ensuring accurate and real-time pricing comparisons forcing these companies to realize and patch their data vulnerabilities.

A Multiplayer Game with Integrated Real Time Messaging System

What I did: Collaborated with a team of developers to develop a multiplayer hangman game with real-time messaging and a lobby system.

How I did it: Via agile development and using modern web technologies like ReactJS, DaisyUI, and Tailwind CSS for the frontend, and Express JS and Socket.io for the backend.

A Machine Learning Powered Automobile Price Predictor

What I did: Developed a price prediction application for automobiles leveraging various machine learning models such as Linear Regression, Gradient Boosting Regressor, Support Vector Regressor, Decision Tree Regressor, and Random Forest Regressor with different normalization methods such as PCA and K-fold validation.

How I did it: Performed comparison and analysis to gain insights on metrics like accuracy, F1 score, and confusion matrix using Statistics and Regression Analysis and utilizing concepts like Principal Component Analysis and Linear Discriminant Analysis.

A Virtual Juke Box Application

What I did: Collaborated on a team of 7 members to design a music application for the 2023 BeachHacks hackathon that can help entertain large crowds by allowing users to add their choice of songs into a single collaborative queue.

How I did it: Built reusable front-end components such as Now Playing, Navigation Bar, Queue system, and Search functionality using JavaScript, ReactJS, ExpressJS, MaterialUI, and various available Spotify APIs.

A Meeting Place Finder Application

What I did: Developed a mobile application that finds feasible meeting places between two locations using the user’s input location and the user’s present location. Ultimately, providing the user with the top ten possible locations along with options to filter the options based on the user’s preferences.

How I did it: Front-end components for Android and iOS were built using the React-Native framework, and TomTom’s Search and Places API were used to get the reviews and photographs of the locations.