BKSBhupender Kumar Saini
Research focusCV

Aug 2023 – Nov 2024/Industry Project

Automatic Tender Document Processing and Offer Generation

  • Industry Project
  • Industry Project with FEGA & Schmitt Elektrogroßhandel GmbH
Partner
FEGA & Schmitt Elektrogroßhandel GmbH
Period
Aug 2023 – Nov 2024
Focus
Industry Project

Key outcomeAccuracy@3 of 95 percent and Precision@3 of 90 percent: a relevant product appeared among the first three retrieved results in 95 percent of evaluated cases, at 90 percent precision

System architecture diagram showing raw data flowing through data preparation, pretraining and fine-tuning, a model repository, a retriever with an article database, and an evaluation stage feeding a webapp
End-to-end architecture: data preparation and model fine-tuning through to retrieval, evaluation, and the delivered webapp.

Context

Tender documents and product catalogs often describe the same technical requirement in different words, so keyword search misses relevant products even when a good match exists.

Inputs

  • Tender documents in PDF and text formats
  • Product catalogs and internal databases

Approach

  • Designed a semantic search and retrieval system using LLMs and vector databases
  • Implemented document parsing, metadata extraction, and ranking strategies
  • Built API-based integration components toward deployment in partner infrastructure

Outcomes

  • Accuracy@3 of 95 percent and Precision@3 of 90 percent: a relevant product appeared among the first three retrieved results in 95 percent of evaluated cases, at 90 percent precision
  • Faster search and offer preparation compared with manual lookup
  • Deployed into production on Databricks, followed by code handover and a knowledge-transfer workshop
  • Enterprise search is rarely just about matching words; it requires understanding how customers, engineers, and catalogs describe the same requirement differently

Related on the shelf