Sneh Chavan

Project: AI Lens – Intelligent Document & Image Q&A EngineRole: Lead AI Product Engineer (End-to-End Development) Tools & Tech: Python, LangChain, Gemini Flash/OpenAI API, Tesseract OCR, ChromaDB, Streamlit.Project Overview: Designed and built a multi-modal RAG (Retrieval-Augmented Generation) system that enables users to "talk" to unstructured data sources, including complex PDFs, scanned documents, and images. The project focused on solving the problem of data silos in non-digital formats by creating a pipeline that extracts, embeds, and queries information with high semantic accuracy.Key Achievements:Multi-Modal RAG Pipeline: Developed a modular architecture using LangChain and ChromaDB to manage high-density document ingestion. Implemented recursive character splitting to ensure that context was preserved across large datasets.Hybrid OCR Extraction: Integrated Tesseract OCR and OpenCV to preprocess and extract text from low-fidelity scans and handwritten notes, making "invisible" data searchable for the first time.Vision-Language Integration: Leveraged Gemini Pro Vision to enable the system to interpret non-textual data, such as charts, diagrams, and logos, providing a comprehensive understanding of the document beyond just the text.Contextual Filtering & Accuracy: Fine-tuned prompt engineering strategies to reduce hallucinations, ensuring that the system provided citations and referenced specific page/image regions for its answers.Rapid UI Prototyping: Built a production-ready interface using Streamlit, allowing for real-time file uploading, vectorization, and interactive Q&A, taking the project from concept to functional MVP in record time.

Project: AI-Receipt-Accountant – Automated Financial Intelligence EngineRole: Lead AI Product Engineer (End-to-End Development)Tools & Tech: Python, Gemini 1.5 Flash / GPT-4o, Streamlit, Pillow (PIL), JSON Schema, RegEx.Project Overview:Developed a specialized computer vision and NLP pipeline designed to automate expense management and financial auditing. The system transforms unstructured images of physical receipts and invoices into high-precision, structured financial data, eliminating the need for manual data entry and reducing human error in accounting workflows.Key Achievements:Structured Field Extraction: Engineered advanced prompt templates and JSON Schema validation to force LLM outputs into strict, machine-readable formats. This allows for seamless integration into downstream databases or CSV exports.Vision-Language Preprocessing: Implemented image handling using Pillow to manage varied photo qualities, orientations, and lighting conditions, ensuring high "Read-Rates" for even low-quality or wrinkled receipt captures.Automated Taxonomy Categorization: Built an intelligent classification layer that maps merchant names and itemized lists to specific expense categories (e.g., Travel, Logistics, Supplies) using contextual analysis.Multi-Model Verification Logic: Designed a comparison workflow that utilizes multiple LLM APIs (Gemini/OpenAI) to cross-verify critical financial fields like "Tax Amount" and "Grand Total," ensuring professional-grade data integrity.Production-Focused Interface: Developed a high-speed Streamlit dashboard that supports bulk uploads and real-time data editing, providing a "one-click" experience for generating financial reports.

Project: Drone Sentinel AI 2.0 (Security Orchestration Engine)Role: Lead AI Product Engineer (System Architect)Tools & Tech: Python, LangChain, YOLOv8, FastAPI, Geo-spatial Algorithms, Docker.Project Overview: Developed an autonomous AI orchestration engine designed to transform raw drone telemetry and visual data into structured security intelligence. The system serves as a central "Command and Control" (C2) layer, performing real-time threat evaluation and flight path deconfliction for drone fleets in sensitive airspace.Key Achievements:4D Threat Deconfliction: Engineered a proprietary spatial-temporal algorithm to predict and prevent mid-air collisions and security breaches by analyzing flight waypoints in 3D space plus time.Automated Intelligence Pipelines: Integrated YOLOv8 for real-time object detection, allowing the system to autonomously identify and categorize unauthorized entities within a secured perimeter.System Orchestration: Built a modular API-driven backend using FastAPI to manage high-frequency data ingestion from multiple drone sensors, ensuring sub-second latency in decision-making.Conflict Resolution Engine: Designed a "Conflict Explanation" module that provides human-readable logs and actionable alerts, moving beyond simple "detected" statuses to detailed "why and where" security reports.Production-Ready Deployment: Containerized the entire stack using Docker, ensuring that the orchestration engine could be deployed consistently across various edge and cloud environments.