Laxman Madasu

Developed an MCQ Generator Web Application leveraging Streamlit for an intuitive UI and the Gemini model for instant, high-quality question generation. This tool empowers users to transform any text into customizable multiple-choice questions (5, 10, 15, or 20), streamlining content assessment and learning material creation. It provides an efficient solution for educators and trainers to rapidly produce engaging MCQs.

Developed a Python application using Streamlit and OpenAI API to review code and provide feedback on bugs and fixes. Created a simple interface where users can submit their code and get instant feedback. Implemented an efficient system to analyze code, detect bugs, and suggest accurate fixes using the OpenAI API. Made the tool easy to use, helping developers improve their code quickly.

Engineered a Convolutional Neural Network (CNN) model to accurately predict dog breeds from images using a comprehensive dataset. Created an interactive Streamlit application that allows users to upload dog images and receive real-time breed predictions. Implemented functionality to visualize and display extracted features from the CNN at every convolutional layer, enhancing model interpretability. Optimized model performance through techniques such as data augmentation and transfer learning.

Evaluated hotel reviews for sentiment classification using machine learning techniques. Preprocessed text data and extracted features with BoW, and TF-IDF. Trained and evaluated models: Naive Bayes, Logistic Regression, and XGboost. Developed and deployed a real-time sentiment analysis app with streamlit. Achieved 83% accuracy in sentiment classification and documented the project on GitHub.

Built an AI-powered smart scanner for a finance application that extracts structured information from receipts and financial documents and stores processed data in MongoDB. Developed a Streamlit Smart Scanner that allows users to scan/upload receipts, automatically extract relevant fields, and populate the extracted information into a structured form for user review and correction. Implemented an OCR-based document processing pipeline using PaddleOCR, Tesseract, PyMuPDF, and pdfplumber to extract text and map document content to predefined fields. Integrated LLMs with document processing workflows to handle complex financial documents and convert unstructured document data into structured JSON suitable for downstream processing. Designed the workflow to persist validated extracted information into MongoDB, enabling users to review, update, and manage processed financial records. Developed transaction classification capabilities combining rule-based logic and fuzzy matching to categorize financial transactions and improve classification using user feedback.

Engineered a FastAPI + MongoDB LLM platform converting natural-language prompts into schema-valid UI page trees for websites, portfolios, and reusable components. Architected a multi-stage RAG generation pipeline (planner → grounded generator → validator → self-repair → quality scorer) that lifted structure validity from 33% to 100% and component correctness from 98.8% to 100% across a 12-page evaluation set. Implemented hybrid RAG retrieval (dense embeddings + BM25, fused via Reciprocal Rank Fusion) over a 154-component knowledge base, eliminating hallucinated component types from generated output. Built a provider-agnostic LLM layer spanning 7 providers (Gemini, Groq, OpenAI, Anthropic, Cohere, DeepSeek, Ollama) with cross-provider fallback and per-stage token/cost accounting. Delivered SSE streaming generation and session version history (undo/revert), validated by a 300-prompt benchmark corpus and 265 unit tests.

Developed a multi-tenant LLM application platform enabling non-engineers to build and publish AI applications across three execution modes: conversational chat, tool-calling agents, and deterministic flows. Designed a provider-agnostic LLM abstraction across 7 providers (OpenAI, Anthropic, Gemini, Groq, Cohere, DeepSeek, Ollama), letting each application swap models purely through configuration. Established production reliability primitives from scratch — circuit breakers, MongoDB-backed distributed locking, idempotency gating, dead-letter queues, and retry/timeout policies — validated by a 190-test suite. Enforced strict multi-tenant data isolation via a database-proxy layer injecting tenant context into every query, combined with deny-by-default tool scoping to block cross-tenant access. Diagnosed and resolved a critical multi-turn agent reliability issue (tool-call success stuck at 57%) by fixing cross-turn context persistence, restoring reliable multi-step tool execution.