projects
Academic Projects
Multi-Agent LLM System for Automated App Development
Aug 2025 – Oct 2025- LangGraph Pipeline: Built a planner → architect → coder pipeline that turns plain-language specifications into runnable project scaffolds with controlled recursion.
-
Deterministic Task Execution: Integrated Groq’s
gpt-oss-120bvia LangChain with Pydantic structured outputs to keep task plans deterministic. - Sandboxed Tool Layer: Hardened tool layers for sandboxed file and command access, enabling safe automated generation of multi-file web app prototypes.
AI-Powered Knowledge Graph Builder
Jan 2025 – Apr 2025- Automatic Entity Extraction: Built a knowledge-graph backend combining FastAPI with Ollama and OpenAI LLM pipelines for automatic entity and relationship extraction from unstructured documents.
- Production Ingestion: Engineered secure ingestion with MIME validation, path traversal defense, and suspicious content scanning with version-controlled graph history and REST APIs.
- Database Persistence: Optimized persistence via SQLAlchemy models, Alembic migrations, and batched database operations for low-latency queries.
Intelligent Cyber Intrusion Detection System
Aug 2023 – Dec 2023- Feature Engineering: Developed a deep learning-based IDS using the NSL-KDD dataset, optimizing feature selection with Recursive Feature Elimination (RFE) and a Decision Tree classifier.
- CNN-LSTM Model: Achieved 95% accuracy and 0.94 F1-score with the CNN-LSTM model, demonstrating superior performance in intrusion detection.
- Publication: Research accepted for publication in Elsevier Computers & Security (Q1, IF: 6.8).
E-Commerce Product Recommendation System
Dec 2021 – Feb 2022- Feature Extraction: Encoded product images into dense feature vectors using pretrained ResNet50 backbone with Global Max Pooling.
- Visual Search: Applied L2 normalization and brute-force k-Nearest Neighbors (Euclidean distance) search to rank visually similar catalog items.
- Real-Time Interface: Created a Streamlit interface that uploads customer images, infers embeddings on the fly, and displays top recommendations in real time.