I'm Srinjoy Ghosh, a Data Scientist and AI/ML Engineer with a Master's in Data Science from Friedrich-Alexander-Universität Erlangen-Nürnberg, specializing in Artificial Intelligence, Machine Learning, and Software Development.

• Software Development and Testing and AI/DS Consulting at RX-Systems, Berlin, Germany
• AI/ML Engineer (HIWI) at RX-Systems, Berlin, Germany
• Data Scientist (HIWI) at Fraunhofer IISB, Erlangen, Germany
• Python Django Trainee at Golden Eagle IT Technology, Kolkata, India
• Data Analyst Intern at Ausavi AI Pvt. Ltd., Bangalore, India

• Designed and implemented RICHES (RAG with Intelligent Cross-Document Clustering and Hybrid Ensemble Search), a novel self-hosted multilingual Retrieval-Augmented Generation system for secure document processing across English and German.
• Developed a novel multimodal PDF parser achieving state-of-the-art extraction of text, tables, images, and mathematical equations using layout detection (DETR) and specialized OCR, operating entirely locally without VLM dependencies.
• Engineered a three-pipeline architecture (Ingestor, Retriever, Generator) with progressive multi-strategy retrieval: Simple Search for factual queries, Chain-of-Thought for comparative analysis, and Agentic Planning for complex multi-step inquiries.
• Implemented ingestion-time K-Means clustering with Elbow Method optimization, transforming brute-force search into efficient two-stage lookup (cluster centroid → members), drastically reducing query latency while maintaining accuracy.
• Achieved 93.7% completion rate vs. 30% for OpenAI Playground on heterogeneous benchmarks, processing 100 files across PDFs, DOCX, images, and structured data with specialized parsers within 24GB VRAM constraint.
• Built hybrid parallel retrieval combining vector search, BM25 lexical search, SQL metadata search, and image similarity using Reciprocal Rank Fusion (RRF), with language-consistent generation and inline citation traceability.
• Processed complete ingestion pipeline in 10m 46s with predictable cost distribution and sub-minute query response times, demonstrating 3.1× better search accuracy than commercial alternatives.