Tanuj Kumar Verma

Project Title: Ask Marcel (Agentic Langgraph)

Technologies: FastAPI, LangGraph, LangChain, Azure OpenAI (GPT + Embeddings), Azure Cognitive Search (Semantic + Vector), Redis (async), OpenTelemetry + Azure Application Insights, PostgreSQL/Azure SQL (SQLAlchemy/asyncpg/pyodbc), GraphQL (Neo4j), tiktoken.

Description: Ask Marcel is an agentic GenAI service built on LangGraph, orchestrating specialized agents (supervisor, RAG, small‑talk, summarizer, database‑operation) to answer employee questions in real time. FastAPI exposes a streaming SSE chat API secured by JWT; Redis powers low‑latency session, intent, and embedding caches. Retrieval fuses Azure Cognitive Search (semantic + vector) with Azure OpenAI embeddings/LLMs and LangChain adapters to refine queries, rank sources, and craft grounded answers. User metadata and profile data (via PostgreSQL/Azure SQL and Neo4j GraphQL) enable personalization and community‑aware filtering. OpenTelemetry + Azure Application Insights deliver distributed tracing and logs. Production hardening includes CORS, robust error handling, connection pooling, async IO (httpx/aiohttp), and tiktoken‑based prompt budgeting.

Project Title: Microsoft Azure AI Search

Technology: Microsoft Azure Services, Azure AI Search, Azure OpenAI.

Description: I am developing a platform using Microsoft Azure AI Search to interact with documents stored in Azure Blob Storage and SharePoint. The platform creates indexes and uses skillsets to enhance response accuracy. It integrates with Azure OpenAI for query processing, providing a chat playground for user interaction. A web application has been deployed for organizational users to test the combined functionalities of Azure AI Search and OpenAI.

Project Title: Small Language Models.

Technology: Small Language Models (SLM), RAG (Retrieval Augmented) Generation, CTransformers, LlamaCPP, Query Engines.

Description: Conducted extensive research on small language models that can run on CPUs without requiring GPUs. Developed a RAG application in order to analyze accuracy and the period of time acquired by disparate small language models to give responses while running on CPUs rather than GPUs.

Project Title: CV Parser Tool (Generative AI)

Technology: Langchain, LlamaIndex, Large Language Models, WinSCP, PuTTy, Pageant, Apache, NginX, EC2 Server, Python, RAG (Retrieval Augmented Generation), HuggingFaceHub, OpenAI, MySQL, GitLab, Azure Services.

Description: Developed an online CV processing system, utilizing Azure. Services and machine learning models for efficient document analysis and user query handling. Implemented an offline CV evaluation system using Streamlit, integrating advanced ML algorithms to process multiple CVs concurrently for comprehensive applicant analysis and automation.decision-making. Designed a user-friendly interface for efficient management of applicant data and streamlined evaluation workflows. Deployed the CV parser model on an EC2 server via SCP, integrating GitLab for version control and secure code management. Utilized CI/CD pipelines between GitLab and EC2 for automated updates, enhancing operational efficiency and reliability

Project Title: AI-Driven Document and Invoice Processing with Intelligent Data Extraction

Technologies: Python, YoloV5, PyTesseract, Regex, Excel, LabelImg, Makesense.ai

Description: Developed and implemented AI-driven solutions for document and invoice processing. Built and trained YoloV5 models for image labeling, specifically for identity and non-identity cards, ensuring high accuracy on real-time datasets. Designed an Address Parser to process Australian addresses, breaking them into structured components such as Street, City, State, Country, and Pincode. Implemented an Invoice Parsing System using PyTesseract and Regex to extract and categorize entities including Invoice Number, Invoice Date, Amount, and Sender/Receiver details. Additionally, managed large-scale invoice coordination at Alkem Laboratories by automating data extraction, handling rejected invoices, maintaining detailed Excel-based records, and conducting product matching to ensure precision in medicine name identification during production processes.