Chirag
Sharma

Built a conversational AI agent that answers personal finance queries and performs Indian tax calculations from natural language input, deployed live on Streamlit.

• RAG-based document retrieval for financial policy Q&A with source grounding by using Indian Budget 2026 highlights and Post Office Savings scheme data as the knowledge base
• Browser-based tax computation via Playwright for real-time official calculation
• Used MiniLM-L6 (HuggingFace) for semantic embeddings and LLaMA-3.1-8B (GroqAPI) as the backbone LLM.
• Exposed functionality via a REST-style API with persistent conversational memory; orchestrated with LangChain and instrumented with Python logging for backend observability.

Developed an end-to-end anomaly detection pipeline over ~23 million maritime AIS signals, replacing a static rule-based system with a dynamic, AI-driven workflow.

• End-to-end pipeline over 6 months of per minute resolution (~ 250,000 rows × 90 features) maritime sensor data
• Implemented and benchmarked 3 Machine Learning and Deep Learning models — Isolation Forest, LSTM Autoencoder, and Transformer Autoencoder — using reconstruction-error-based anomaly scoring; enabling data-driven threshold tuning.
• Engineered root-cause analysis module to surface the top contributing features for each detected anomaly, giving operations teams quantitative evidence for prioritizing maintenance.
• Reduced data load latency by 98% (from ~2 min to under 2 sec) via Parquet-based optimization, enabling rapid exploratory analysis and real-time reporting.

Project Associate at Ahmedabad University with ISRO's Space Applications Centre. Comparative study of filament detection algorithms on astronomical datasets to identify which best detects interstellar filamentary structures from observational data. Our findings suggested that DisPerSE and FilFinder tend to be more generous in detecting filaments, while getSF tends to be more conservative and detects fewer, dominant filaments. On the quantitative side, we found that MSSIM (Mean Structural Similarity) index was not a good metric for comparing the performance of the algorithms as it did not behave as expected (Green et al. 2017). We concluded that new metrics are needed to compare the performance of the algorithms in an unbiased, objective manner.

• Researched and benchmarked multiple interstellar filament detection algorithms (FilFinder, getSF, DisPerSE) across diverse astronomical datasets, contributing to pipeline selection for ISRO's imaging analysis workflows.
• Quantitatively evaluated algorithm performance using the MSSIM structural similarity index, enabling objective comparison across varied observational conditions.
• Presented research findings at the SAMHITA-2025 national conference hosted at SAC, ISRO, Ahmedabad

Master's thesis at IISER Bhopal. Developed custom numerical solvers for the coupled PDEs arising in linear cosmological perturbation theory — the mathematical framework for understanding large-scale structure formation in the universe. A good reference is Callin 2006.

• Solved full coupled PDE system from linear cosmological perturbation theory
• Implemented 4th-order Runge–Kutta (RK4) for time evolution (unstable!)
• Implemented adaptive RK5 with Cash–Karp parameters for improved stability
• Results interpreted in context of CMB anisotropies and structure formation