The Challenge
During my time at Bharati Innovative, I encountered a common engineering challenge: valuable historical design knowledge was scattered across disconnected Google Drive files. Staff attrition and the lack of a centralized search system meant engineers often spent significant time looking for past solutions or unintentionally re-inventing the wheel during new part design.
What I Built
To address this, I built a Proof of Concept (POC) for an AI-driven Retrieval-Augmented Generation (RAG) system to organize and unlock the company’s internal knowledge base. The system used intelligent routing to determine whether queries could be answered directly by an LLM or required real-time company data, while document chunking, embeddings, and Pinecone enabled fast semantic search. I also deployed n8n with Docker and used Ollama to run open-source LLMs locally, maintaining data privacy and tailoring outputs for CAD/CAM engineers and toolmakers.
- Intelligent Query Router: Distinguishes general domain questions from proprietary engineering lookups requiring real-time internal technical context.
- Semantic Ingestion & Chunking: Pipelines that extract and chunk CAD/CAM design notes and historical solutions into rich vector representations.
- Pinecone Vector Database: High-performance index for low-latency similarity queries across thousands of historical engineering documents.
- Local LLMs via Ollama: Runs privacy-first open-source models on-premise without exposing sensitive IP or proprietary blueprints to external APIs.
- Workflow Automation with n8n & Docker: Containerized microservices pipeline that standardizes formatting, handles CAD/CAM terminologies, and delivers contextual answers.
Impact & Results
By making historical technical solutions instantly searchable, the system reduced design-phase lookup time, minimized redundant work, and improved overall workflow efficiency by an estimated 65–75%.
Huge thanks to my mentors and the team at Bharati Innovative for their guidance and support throughout the project!