Srinath Suresh

Computer Science professional with a strong foundation in AI/ML, computer vision, and Natural Language Processing. Experienced in generating data-driven insights, developing innovative solutions, and demonstrating a strong ability to collaborate in team environments. Committed to delivering quality results while maintaining a focus on continuous learning.

• Anomaly Detection using Advanced Snort | Python, Scikit-learn, Snort, Built a hybrid intrusion detection workflow combining Snort rule-based detection with machine learning to classify network traffic as normal or anomaly., Trained and compared multiple ML classifiers, including Random Forest, Decision Tree, SVM, and KNN, using network packet features to detect anomalous traffic., Achieved 98.99% accuracy with Random Forest, as the best-performing model, and explored Python based Snort Rule updates to respond to suspicious network traffic.
• Multi-Agent LLM Framework for Logical Puzzle Solving | Python, Gemini API, Groq API, Designed a multi-agent LLM reasoning framework to solve grid-based logic puzzles by coordinating clue interpretations across multiple agents., Modelled inter-category relationships using probability matrices to track possible entity matches and iteratively refine puzzle solutions., Integrated Gemini and Groq APIs to support distributed agent collaboration and compare reasoning outputs across different LLM responses., Evaluated the framework on the Puzzle-Eval benchmark, correctly solving 45 out of 276 puzzles without hard-coded rules while identifying scalability limits in long-context reasoning.
• Kimono Color Recommendation System | Python, BeautifulSoup, Scikit-learn, K-means, PCA, Built an unsupervised color recommendation system to generate harmonious kimono-inspired color palettes from 4,524 scraped kimono images., Extracted OKLab-based 512-dimensional histograms using center cropping and pixel sampling, then applied StandardScaler, PCA, and K-means++ to reduce features from 512 to 50 dimensions and group similar color palettes., Implemented item-based collaborative filtering with cosine similarity to recommend compatible colors, improving LLM-evaluated palette scores from 7–8 to 8–9.
• Agentic Text-to-SQL Framework | Python, SQL, LLMs, Streamlit, BIRD Benchmark, Built an agentic Text-to-SQL framework to translate natural language questions into executable SQL queries for real-world relational databases., Prepared the BIRD Mini-Dev benchmark by extracting schema metadata, table-column relationships, and foreign keys to support SQL query generation., Supported a multi-stage pipeline with schema linking, UnJoin schema simplification, SQL generation, and Judge Node self-correction to improve SQL reliability., Evaluated the pipeline on 500 BIRD Mini-Dev queries across 11 databases, achieving 96%+ execution success through iterative SQL validation and correction.