AI and LLM Systems | Retrieval-Augmented Generation | Applied Machine Learning
LinkedIn: https://www.linkedin.com/in/sudarshanan-santharam-9a596a192/
Portfolio: https://sudarshanangss.github.io/Personal-Website/
I am a Master of IT (Intelligent Systems) student at RMIT University with three years of professional software engineering experience.
My work focuses on designing and implementing AI systems that integrate retrieval pipelines, large language models, and scalable backend services. I am particularly interested in end to end systems that transform unstructured data into structured and actionable outputs.
Key areas of interest:
- Retrieval-Augmented Generation (RAG)
- Hybrid search using BM25 and FAISS
- Multi-agent LLM orchestration
- Multimodal deep learning
- Reinforcement learning systems
- Reproducible experimentation and evaluation
Repository: https://github.com/SudarshananGSS/multi-agent-llm-pipeline
Designed and implemented a modular multi-agent architecture for structured knowledge extraction.
Integrated hybrid BM25 and FAISS retrieval with LLM reranking and stance classification.
Applied RRF, MMR, and RM3 for retrieval optimisation.
Conducted systematic ablation studies with a final Jensen-Shannon score of 0.308.
Repository: https://github.com/SudarshananGSS/multimodal-visual-entailment
Built a multimodal model combining vision and text representations for entailment classification.
Implemented reproducible data pipelines and structured evaluation using classification metrics.
Explored model interpretability using SHAP.
Repository: https://github.com/SudarshananGSS/information-retrieval-pipeline
Designed and implemented a full information retrieval pipeline including indexing, preprocessing, and ranking. Implemented term-based retrieval using inverted indices and BM25 scoring. Integrated query expansion and ranking optimisation techniques to improve retrieval effectiveness. Evaluated performance using standard IR metrics such as MAP and precision at k.
Repository: https://github.com/SudarshananGSS/packet-scheduling-rl
Developed a reinforcement learning scheduler for QoS-aware packet queues.
Designed a custom simulation environment and evaluated latency and throughput trade-offs.
Repository: https://github.com/SudarshananGSS/white-blood-cell-classification
Implemented a multi-head convolutional neural network to classify WBC type and morphological features.
Applied transfer learning, class-balanced sampling, and loss weighting for multi-task optimisation.
- Regression Model for Chatbot Performance Prediction - Regression-based analysis to predict chatbot performance metrics.
- Automated Wine Quality Prediction and Clustering - ML pipeline for wine quality prediction and unsupervised clustering.
- Event Ticket Booking App - JavaFX desktop app for event ticket booking and management.
- Movie Recommendation System (kNN + Matrix Factorization) - Recommender system using collaborative filtering and latent factor models.
- School-age Connectivity and Inclusive Infrastructure Mapping - Geospatial and data-driven analysis of school-age connectivity and infrastructure access.