Projects

Home     About    Skills    Projects    Experience

Project Highlights

Private RAG Pipeline with FastAPI, Qdrant, and Fine-Tuned LLaMA

Document Ingestion (PDF Import Module):

• Implemented a PDF ingestion pipeline that accepts multiple documents through a drag-and-drop UI (Streamlit) or REST API (FastAPI).
• Each PDF is parsed using PyMuPDF or pdfplumber to extract clean, structured text.
• Metadata (e.g., title, source, page count) is captured for later filtering and display.

Embedding Generation & Storage:

• Extracted text chunks (based on token limits) are passed to a selected embedding model (e.g., all-MiniLM-L6-v2 or bge-base-en) using HuggingFace or SentenceTransformers.
• Each chunk is vectorized and stored in Qdrant, a high-speed vector search engine, along with its metadata and original content.
• The vector index is organized by document and optionally namespace for multi-tenant support.

Predictive Tagging (Keyword Extraction & Classification):

• For each chunk, we used KeyBERT or a transformer-based classifier to extract semantic tags.
• Tags are added to the Qdrant metadata fields for advanced filtering and smart querying (e.g., “find security-related documents tagged with ‘OAuth’”).
• An option for human-in-the-loop review of auto-tags is included in the Streamlit UI.

Backend Search API (FastAPI with RAG Logic):

• A RESTful API is built using FastAPI, exposing endpoints for:
• Document upload and indexing
• Vector similarity search with hybrid (tag + semantic) filtering
• Prompt-based document querying (RAG)
• API handles chunked retrieval from Qdrant based on user queries, re-ranks results, and composes context windows for LLM input.

Local LLM Integration (LLaMA + Ollama):

• Deployed a fine-tuned version of LLaMA 3 using LoRA/QLoRA for specialization in documentation-based search tasks.
• The model is served locally via Ollama, exposing a simple HTTP API.
• Input prompt includes top-k document chunks retrieved from Qdrant + user query → model generates the final answer or summary.
• Training used a custom instruction dataset based on the ingested PDF corpus.

Streamlit Frontend UI:

• Developed a clean, user-friendly interface to:
• Upload and preview PDF documents
• View automatically extracted tags
• Enter search queries and view generated answers
• Inspect which chunks were retrieved and used in the generation
• UI is directly connected to FastAPI backend via HTTP calls

Real-Time Analytics Platform

• Objective:
  Developed a real-time analytics platform to process and analyze data for business intelligence applications.

• Technologies Used:
  Debezium, Kafka, Python, Greenplum, Docker, Kubernetes, Graylog, Prometheus, Grafana

• Key Achievements:

  • Achieved sub-second latency in capturing and processing database changes
  • Designed and implemented scalable microservices capable of handling millions of events daily
  • Built Grafana dashboards to monitor system performance and critical operational metrics

Microservices-Based Data Pipeline

• Objective:
  Developed a microservices-based architecture to enable real-time streaming and processing of large datasets from multiple relational databases.

• Technologies Used:
  Debezium, Kafka, Python, Docker, Kubernetes

• Key Achievements:

  • Automated deployment and scaling of services using Kubernetes
  • Reduced operational complexity by centralizing logs through Graylog containers
  • Improved fault tolerance via Kafka partitioning and replication strategies

GPS Tracking & Telemetry Data Pipeline

• Designed and maintained high-performance, real-time data pipelines focused on GPS tracking systems and telemetry data ingestion.
• Engineered solutions capable of processing hundreds of millions of records per day to support real-time analytics.
• Leveraged cloud infrastructure, containerization, and data streaming technologies to optimize workflows for high-velocity, high-volume environments.