Pooja Shinde

Client – Afcons Infrastructure
Project – AI Knowledge Management System (Agentic AI + KAG + Multimodal)

• Built an end-to-end multimodal Knowledge Management System to process construction data (documents, images, audio, video) using LLaVA (image/video captioning), Whisper (speech-to-text), and PaddleOCR (text extraction) pipelines.
• Developed an Agentic AI system using Google ADK, enabling natural language querying over Knowledge Graphs via LLM-generated Cypher queries and Neo4j-based graph retrieval.
• Implemented Knowledge-Augmented Generation (KAG) using Microsoft GraphRAG + Neo4j, with Azure OpenAI (GPT-4o) for grounded answer generation and citation-based responses.
• Fine-tuned a LLaMA 3.x (~3B) model using multi-task instruction tuning (Q&A, summarization, prompt refinement) on domain-specific Supply Chain Finance data to improve reasoning and domain accuracy.

• Worked on early-stage GenAI solutions (2023–2024), including:
  Developed a Text-to-SQL system using LLMs to convert natural language queries into SQL for business users.
  Implemented a Retrieval-Augmented Generation (RAG) pipeline to extract insights from structured and unstructured financial data.
• Developed and deployed machine learning-based credit risk models using Logistic Regression, XGBoost, and Random Forest to improve lending risk assessment.
• Built automated and scalable ML pipelines for data preprocessing, feature engineering, and model deployment