Amit Kumar Gupta

A powerful retrieval-augmented generation (RAG) based question-answering system designed to help users explore their own PDF data, maybe containing proprietary or the latest data, using LLMs with context-aware responses. The system is capable of utilizing multiple LLMs (local via Ollama and remote ones, such as OpenAI and Gemini). It uses Qdrant DB to store vector embeddings.

I created a completely offline RAG system in this project. We can upload any type of document, such as PDF, HTML, etc., to our system. A knowledge system is built by first using Docling to parse the documents into text and then saving that info in LanceDB as a vector store. Then, it can answer any user query by using the knowledge system and Gemma3 LLM. Everything (including the vector store and LLM) is running locally, so there's total privacy and free operation.

This project is a coding-focused AI assistant designed to help users solve programming problems and answer coding-related queries. The assistant follows a structured, step-by-step process—Start, Plan, Action, Observe, and Output—to ensure clear and logical problem-solving. It can execute terminal commands securely and provide detailed explanations for each step in the process.

• Fine-tuned a pre-trained Flux.ai diffusion model to generate images of a specific person and style using only 25–30 images, without retraining the full model.
• Performed dataset curation and captioning, using a unique trigger token to condition the diffusion model for subject-specific and style-consistent image generation
• Implemented supervised fine-tuning (SFT) by training LoRA adapters on attention layers while keeping base model weights frozen to reduce compute and memory usage.
• Evaluated model performance by comparing base model outputs vs LoRA-adapted outputs for identity preservation, visual consistency, and prompt adherence.
• Enabled efficient inference and deployment by dynamically loading LoRA weights, allowing modular reuse of the base model for multiple personalization tasks.

I built a production-grade sentiment analysis system capable of processing massive review datasets using memory-efficient HashingVectorizer and incremental learning with SGD Classifier—handling datasets too large for traditional methods.

This project demonstrates how easily we can use an LLM locally. It uses Llama 3.2 vision and Streamlit to create a 100% locally running OCR app. We can give any scanned image to the app, and it will attempt to extract text from it. I demonstrate how to set it up completely locally if you have a decent machine, or use the free hardware provided by Google Colab to still run your code pretty much privately. We set up Ollama in Google Colab and then use LocalTunnel to expose the system to the outside world!

Vendors regularly reach out to our organization's customer care regarding the status of their invoice. An innovative AI-based system was proposed to be developed to automate the resolution of such queries. The project has been initially developed as an Agentic AI system that can answer vendor queries via email. Vendors can send in their invoice-related queries via email, and the system replies to them. The USP of the solution is that vendors can send their queries in standard English (also called 'Natural Language'). The solution leverages Oracle Integration Cloud (OIC) with an Email Adapter to monitor a dedicated inbox, integrates with LLM (OCI Generative AI using Cohere LLM) to extract intent and parameters from the email's natural language query, and then executes the appropriate Oracle Cloud operations through REST APIs. It then responds to the user via email with a natural language-sounding response.

Emerson's internal database team wanted to enable internal business customers to query the database using natural language.

My role:
• Collected the requirements from the customer.
• Developed an OpenAI API-based solution.

The project required the development of NFTs from scratch and then auctioning them off. A smart contract for NFT was developed based on ERC721, and the smart contract for the auction was developed as an upgradeable contract. The NFTs are being used as membership proofs for clients' real estate customers.

• Developed NFT contracts that support ERC2981 and allow custom royalty for each of them, along with a default one. It conformed to the operator filter requirements of OpenSea.
• Split sale and streaming proceeds with creators using a splitter contract. Researched various protocols like OpenZeppelin's PaymentSplitter and 0xSplits and chose the best one suited for the project's requirements.
• Harnessed OpenSea API to list NFTs programmatically.
• Distributed initial sale amounts among creators. Used Ethers.js queryFilter to connect to blockchain using WebSocket connections and get all transfer events. Used Etherscan APIs to scan internal transactions for OpenSea payments.
• Created a React-based UI so users can log in. For admin users, the UI lists all the minted NFTs of their current owners. The UI allows the admin to interact with the NFT and splitter contracts to distribute funds.
• Used solidity-coverage for test coverage, hardhat-gas-reporter to analyze and optimize contracts and methods, and Slither for vulnerability analysis.

As part of federal requirements, companies are required to maintain data digitally. However, many documents like invoices and purchase orders, while in digital form like a PDF or a text document, are not searchable. Not much value can be extracted from them. Emerson engaged an industry-leading company called Esker, which is capable of scanning the invoice files in various formats (PDF, ZIP, XML) and then identifying the different elements and allowing Emerson to feed this into a data warehouse that stores each element in a searchable manner.

This multi-year, ongoing project is executed by one ERP system at a time.

MY ROLE
• Worked on multiple projects and connected two Oracle ERP and seven SAP instances to Esker.
• Used MOVEit and OIC to deliver files and receive acknowledgments. The project involved SFTP transfers as well as REST API-based deliveries.

The integration center of excellence team was to provide a central solution that all Emerson verticals could harness to exchange files within and outside Emerson. The solution needed to be flexible, scalable, resilient (automatic and manual re-submission), easy to monitor, audit, secure, and maintainable.

MY ROLE
• Acted as the tech lead for this project.
• Implemented Oracle SOA Suite-based architecture to pick data from sources like SFTP, FTPS, and HCM Cloud, processing the files (encryption, decryption, zip, rename) and delivering to varied targets.
• Automated code development so developers could generate code using a Java utility.

The agent can display a list of courses offered by an educational institute after it has been fed that information via PDF documents. The agent was developed using D-ID and is surprisingly fast to develop.

• Formulated a manufacturing plant producing standard and premium products under shared resource constraints.
• Implemented binary setup variables, capacity constraints, and profit maximization logic using Gurobi’s Python API.
• Compared continuous LP solutions with integer-restricted solutions to demonstrate operational feasibility.