×
0 followers0 following
Devashish Kapte
Maharashtra, India0 followers0 following
- About
- High School
- Bachelor's
- Target
Male
Maharashtra, India
Work ExperienceData Engineer (GenAI Developer)
Genpact - Full Time
Oct 3, 2024 - Present · 1 yr 9 mos
Karnataka, India
Worked on a large scale data migration project.
• Led enterprise-scale data migration from legacy systems to Databricks, architecting a robust Data Lake with Delta Lake
integration for scalable analytics.
• Engineered and optimized Silver and Gold layers in the ETL pipeline, driving high-performance data pipeline optimization
and seamless data transformation.
• Migrated 1,500+ tables and refactored 150+ complex Talend workflows into streamlined workflow orchestration, enabling
modular and efficient pipelines.
• Developed DDL scripts for 500+ tables, integrated CI/CD for data delivery, and implemented data quality monitoring with
rigorous unit testing.
• Led a domain-specific team and engaged directly with clients to define and deliver solutions aligned with evolving business
data requirements.
ProjectsInvoice-Data-Extractor-Agent
Role - Generative AI Developer
for Personal
Sep 29, 2025 - Oct 24, 2025
Project overview
This Streamlit app is an LLM-driven invoice processing system (Azure OpenAI + LangChain) that:
Reads PDF invoices and extracts structured invoice information as JSON.
Appends extracted invoice records into a universal table (invoice_table.csv).
Builds a small VectorDB index on the universal table to power a live invoice chatbot.
Provides a per-PDF evaluation tab to compare extraction results (ground truth from a newly uploaded PDF) with the stored universal table (prediction) and compute precision/recall/F1/accuracy.
The agent uses three tools:
ReadPDF — extract raw text from uploaded PDFs (PyPDFLoader).
InvoiceExtractor — call LLM to parse invoice text and return structured JSON.
InvoiceQuery — index the universal table text in a Chroma vector store and answer queries using RetrievalQA.
Key features
Multi-file PDF upload and batch processing.
LLM-based JSON extraction and validation (LLM invoked via LangChain wrapper for Azure OpenAI).
Universal CSV table storing invoices and a download button to export all records.
Live chat agent that can use tools to answer questions about the stored invoices.
Evaluation tab that compares fields between a freshly uploaded PDF and a matched invoice in the universal table and computes classification metrics.
Research PapersDevashish Kapte has not added any research paper yet.
VolunteeringDevashish Kapte has not added any volunteering detail yet.