PDFConverse: RAG Chatbot
The Challenge
Analysts waste countless hours manually searching through large PDF documents, leading to inefficiency and frustration in information retrieval workflows.
The Solution
Multi-RAG Chatbot delivering verifiable, contextual answers using Langchain, FAISS, and OpenAI technologies.
Impact
80%
Time Reduction
Faster information retrieval
PDFConverse: Architecture & MLOps
PDF Processing
Documents chunked and embedded for efficient vector storage
FAISS Vector Store
Optimized retrieval using semantic search capabilities
RAG Query
Context-aware LLM generates accurate responses
Output Validation
Verifiable answers with source citations
Evaluation & Continuous Improvement
Iterative prompt tuning optimizes contextual Q&A without costly LLM fine-tuning. Thumbs up/down feedback curates labeled data for continuous refinement.
- Manages LLM context window limits efficiently
- Optimizes API cost dependencies
- Maintains faithfulness score above 90%
90%
Faithfulness Score
ThreatLens: Cyber Dashboard
Security Challenge
Security analysts overwhelmed by noisy, unprioritized RSS feeds, leading to delayed threat response and missed critical vulnerabilities.
Intelligent Solution
LLM-powered aggregation with automated summarization and objective risk scoring using OpenAI and Streamlit.
50%
Triage Time Reduction
Automated risk scoring
ThreatLens: Intelligence Pipeline
1
RSS Stream Ingestion
Real-time threat feed collection
2
Micro-Batch Processing
Efficient data aggregation
3
OpenAI Scoring
Prompt-engineered risk assessment
4
Dashboard Display
Prioritized threat visualization
Validation & Accuracy
Offline regression testing validates scoring prompts against labeled threat history. Structured JSON output ensures consistency and prevents model hallucination in critical data points.
- Data freshness: articles displayed within 30 minutes
- Risk score accuracy exceeds 85% vs. expert opinion
- Continuous prompt refinement based on feedback
85%
Risk Score Accuracy
Supply Chain Commander: Automated Inventory AI
An end-to-end decision engine that predicts future demand (7-days out) and automatically triggers restocking orders to prevent revenue loss.
How It Works
- Dynamic Forecasting: Uses Facebook Prophet to detect complex seasonality (e.g., high demand for winter jackets in Jan, low in July).
- Closed-Loop Architecture: Unlike static notebooks, this system writes forecasts back to a PostgreSQL database for real-time retrieval.
- Actionable API: A FastAPI microservice calculates reorder quantities and serves alerts to a Streamlit dashboard.
Tech Stack
🛠️ Python • Docker • FastAPI • PostgreSQL • Streamlit