Energy-Aware Task Scheduling
PyTorch, DQN, Cloud Scheduling, Reinforcement Learning
Developed an energy-aware cloud task scheduler using a Deep Q-Network (DQN) to optimize task allocation across virtual machines by jointly minimizing energy consumption while maximizing resource utilization and QoS.
- Engineered a multi-objective reward function incorporating CPU, RAM, memory, disk utilization, uptime, and response time.
- Designed a modular RL pipeline comprising environment simulation, agent, scheduler, replay buffer, and evaluation modules.
- Implemented an adaptive ε-greedy exploration strategy, experience replay, and target network updates.
- Benchmarked the scheduler on workloads ranging from 50–1000 concurrent tasks evaluating Pareto-optimal trade-offs.
Flight Operations Data Engineering Pipeline
Python, Apache Airflow, Snowflake, Docker, OpenSky API
Developed an end-to-end ETL pipeline using Apache Airflow to orchestrate scheduled ingestion of live flight-state data from the OpenSky Network API, with automated retries and 30-minute execution schedules.
- Implemented a Bronze–Silver–Gold Medallion Architecture for raw API responses, schema validation, deduplication, and gold datasets.
- Built modular Python ETL components for high-volume telemetry (aircraft metadata, geospatial coords, velocity, heading).
- Integrated with Snowflake for automated table creation, curated data loading, and query optimization.
- Containerized the complete stack using Docker Compose with configuration-driven DAGs and fault-tolerant execution.
MedAssist: LLM-Based Clinical Intake System
Python, FastAPI, React, Gemini, PubMedBERT, Celery
Built an AI-powered clinical intake platform for symptom assessment, medical report analysis, and medication safety using LLMs and NLP.
- Designed a hybrid AI pipeline combining PubMedBERT embeddings, semantic similarity, and LLM reasoning for drug interaction detection.
- Automated SOAP note generation and OCR-based medical report summarization to provide structured clinical insights.
- Implemented real-time risk scoring, QR-based emergency profiles, and Celery-powered medication reminder workflows.
Messy Mashups: Audio Genre Classification
PyTorch, torchaudio, AST, librosa, AdamW
Built a noise-robust audio classification pipeline using Mel Spectrograms and GPU-based preprocessing for efficient training on music genre data.
- Applied Mixup, SNR-based augmentation, and tempo alignment to improve robustness against noisy and time-shifted audio.
- Scaled model from CNN/ResNet baselines to an Audio Spectrogram Transformer (AST) with self-attention.
- Optimized training with AdamW, Focal Loss, and test-time augmentation (TTA), achieving a 0.964 Macro F1 score.
AI-Powered Offline Disk Analyzer
Python, Flask, React, FAISS, Docker, BGE-M3, SigLIP, YOLO
Built a multimodal offline retrieval system over 1000+ files, supporting 5+ query modes including text, image, face, and object search using BGE-M3 (1024-d) and SigLIP (768-d) embeddings.
- Implemented FAISS-based vector search with hybrid reranking, achieving 85–90% accuracy with <5 s latency.
- Enabled text-to-text, text-to-image, image-to-image, and face/object search using YOLO and ArcFace (512-d).
- Integrated LLM-based Q&A for structured offline data exploration.
Data Analyst Agent
Python, LLMs, Pandas, SQL, Matplotlib, Docker
Built an LLM-powered autonomous data analysis agent capable of ingesting and processing large-scale datasets from CSV files, APIs, S3, and SQL sources.
- Designed a multi-stage agentic pipeline with planner, retrieval, analysis, and visualization modules.
- Implemented automated EDA, feature extraction, and statistical analysis using Pandas and NumPy.
- Enabled natural language querying delivering analysis outputs and visualizations in under 3 minutes.
Virtual TA Assistant
FastAPI, FAISS, RAG, LLMs, all-MiniLM-L6-v2
Built an LLM-powered API to answer course-related queries using retrieval-augmented generation (RAG) over course content and 1000+ Discourse posts.
- Engineered a FAISS-based retrieval pipeline with all-MiniLM-L6-v2 embeddings for semantic search and source grounding.
- Designed system supporting JSON-based queries and base64 file inputs, returning responses with source links in <30s.
- Automated data scraping, chunking, and indexing with date-range filtering for continuous knowledge updates.
Cinema Audience Forecasting
Python, XGBoost, LightGBM, Scikit-learn, Pandas
Developed a machine learning pipeline to forecast movie audience demand using historical box office, release metadata, seasonal trends, and genre-specific features.
- Performed extensive feature engineering (temporal features, categorical encoding, missing value imputation).
- Trained and evaluated multiple regression models (XGBoost, LightGBM, Random Forest) with cross-validation.
- Built an end-to-end forecasting workflow covering data preprocessing, model training, and trend visualization.
Protein Secondary Structure Prediction
PyTorch, PyTorch Lightning, BiLSTM, GRU, TrackIO
Developed an end-to-end sequence-to-sequence deep learning pipeline for predicting protein secondary structures from amino acid sequences, jointly learning Q8 (8-state) and Q3 (3-state) annotations.
- Designed and benchmarked 6 neural architectures (BiRNN, BiLSTM, GRU, Inception-BiLSTM, ResNet-BiLSTM, SE-ResNet-BiLSTM).
- Engineered a modular PyTorch Lightning framework with custom tokenization, dynamic padding, and mixed-precision.
- Optimized token-level prediction using Macro-F1 evaluation and cross-entropy based multi-task learning with TrackIO.
Multi-Task Facial Attribute Prediction
PyTorch, PyTorch Lightning, Vision Transformers (ViT), CNNs
Developed a multi-task computer vision pipeline to jointly predict age (regression) and gender (classification) from facial images using shared feature representations.
- Benchmarked custom CNNs and fine-tuned Vision Transformer (ViT) architectures leveraging transfer learning.
- Built an end-to-end PyTorch Lightning framework with advanced image augmentations and mixed-precision training.
- Optimized joint classification and regression losses with Macro-F1 and normalized RMSE metrics.