AI Engineer

Manish Rathi  • Peerlist

Designed and developed a Conversational Text-to-SQL AI system that enables non-technical users to query financial and sales databases using natural language. The system converts user questions into SQL queries using Large Language Models and retrieves insights from structured databases, enabling faster and more accessible business analytics. Key Features: • Built a natural-language-to-SQL agent using LangChain and GPT-based LLMs • Enabled users to ask finance and sales questions in plain English • Automatically generated optimized SQL queries for structured databases • Implemented query validation and error handling for reliable database access • Added automated visualization using Matplotlib and Seaborn for data insights • Achieved ~95% query success rate in simulated enterprise finance datasets Architecture Highlights: • LLM-based prompt engineering for Text-to-SQL translation • LangChain agent orchestration • SQLite structured financial datasets • Automated analytics and visualization layer Business Impact: This system demonstrates how LLM-powered analytics assistants can democratize data access by allowing business teams to retrieve insights without SQL knowledge. Tech Stack: Python, LangChain, GPT/LLMs, SQL (SQLite), Pandas, Matplotlib, Seaborn

Conversational Text-to-SQL Finance Analyst (L

Financial fraud is one of the most critical challenges in digital banking systems. Even a small percentage of fraudulent transactions can lead to significant financial losses for banks and customers. In this project, I developed an end-to-end machine learning pipeline to detect fraudulent credit card transactions using historical transaction data. The workflow included: • Data preprocessing and feature engineering using Pandas and NumPy • Handling severe class imbalance using SMOTE and undersampling techniques • Exploratory data analysis and correlation analysis to identify fraud patterns • Training multiple machine learning models including Logistic Regression, Random Forest, and XGBoost • Hyperparameter tuning and cross-validation to optimize model performance • Model evaluation using Precision, Recall, F1-score, ROC-AUC, and Confusion Matrix • Visualization of fraud detection insights using Matplotlib and Seaborn The final model achieved strong fraud detection performance while minimizing false positives, which is critical for real-world financial systems. This project demonstrates how machine learning can be applied to detect financial fraud and improve transaction security in banking and fintech systems.

End-to-End Credit Card Fraud Detection Pipeli

• Designed and implemented an enterprise-grade Retrieval-Augmented Generation (RAG) system to enable intelligent querying across financial documents (10K invoices and reports). • Built metadata-aware semantic retrieval using FAISS vector indexing with structured filters (year, month, document type) for precise context selection. • Integrated local LLM inference using Ollama (LLaMA) to ensure data privacy and on-prem enterprise deployment. • Engineered document chunking, embedding pipeline, and retrieval orchestration using LangChain. • Implemented source-grounded answer generation with document traceability to reduce hallucination risk. • Developed an interactive Streamlit UI for real-time finance workflow simulations. • Achieved ~40% reduction in query resolution time in simulated enterprise scenarios.

Enterprise-Grade Retrieval-Augmented Generati

Developed an end-to-end financial sentiment analysis system by fine-tuning transformer-based Large Language Models (LLMs) on financial news datasets. The project focuses on extracting sentiment signals from financial news, earnings reports, and market commentary to support data-driven investment and business intelligence decisions. Key Contributions: • Fine-tuned BERT-based transformer models using Hugging Face Transformers and PyTorch for domain-specific financial sentiment classification. • Built a full NLP pipeline including data preprocessing, tokenization, feature engineering, and model training on financial news datasets. • Implemented training optimization techniques including learning rate scheduling, early stopping, and class balancing. • Evaluated model performance using multiple NLP metrics including Accuracy, Precision, Recall, F1-score, BLEU, and ROUGE. • Achieved 92% sentiment classification accuracy on financial news datasets. • Conducted bias and model robustness checks to ensure reliable predictions across different financial topics. • Designed the solution to simulate real-world financial intelligence use cases such as market trend detection, investment sentiment tracking, and risk monitoring. Tech Stack: Python, PyTorch, Hugging Face Transformers, BERT, NLP, Pandas, Scikit-learn, Matplotlib Business Value: Demonstrates how fine-tuned LLMs can extract actionable insights from unstructured financial text to support investment analysis, trading signals, and enterprise financial intelligence systems.

LLM Fine-Tuning for Financial Sentiment Analy

• Designed and implemented an end-to-end retail sales forecasting system using statistical baselines, machine learning, and deep learning models. • Aggregated raw transactional data into daily and category-level time-series datasets with strict time-aware train/test splits (no data leakage). • Engineered lag-based and calendar features for MLP modeling; implemented sliding window sequence generation for LSTM forecasting. • Developed benchmark models (Naive & Moving Average) to ensure fair performance comparison before applying ML/DL. • Built a PyTorch-based Feedforward Neural Network (MLP) for category-level forecasting using standardized features. • Implemented an LSTM model with MinMax scaling to capture temporal dependencies in total sales forecasting. • Evaluated models using MAE and RMSE; LSTM achieved ~17% improvement over naive baseline for total sales prediction. • Generated performance visualizations, residual analysis, and benchmark reports for business-aligned interpretation. • Designed modular pipeline architecture with model saving, inference, and JSON metric reporting.