Eden Sasson

AI engineering mission within a printing company, aiming to transform business data into actionable predictive models across the entire value chain.

Development of an AI-assisted quote generation engine, based on a RAG (Retrieval-Augmented Generation) pipeline to query past quotes and orders. The model automatically suggests consistent pricing based on media type, quantities, finishes, and customer profile — significantly reducing quoting time for sales teams.

Implementation of an intelligent customer follow-up system: detection of orders awaiting validation, contextualized follow-ups, and proactive alerts on delivery times, generated by the model based on the actual production status.

End-to-end data pipeline architecture: ingestion from ERP/MIS, feature engineering, model serving via API, and delivery into existing business tools (CRM, operational dashboard).

Agentic RAG - Intelligent automation of responses to calls for tenders

Responding to calls for tenders (RFPs) is a critical but time-consuming process, often hindered by scattered knowledge. I designed an end-to-end automation system that transforms complex PDF documents into structured and validated proposals, reducing manual effort by 60%.

Challenges Overcome

Complex Extraction: Intelligent parsing of PDF documents (up to 30 pages) with heterogeneous structures (tables, lists, implicit questions).

Response Reliability: Avoiding LLM hallucinations by forcing a triple search (Internal Knowledge Base, History of Past Responses, Real-time Web Search).

Data Maintainability: Implementing a data lifecycle (RFP Age) to ensure the AI only uses the most recent and validated information.

Technical Expertise Deployed

Agentic Architecture (LangGraph & LangChain): Implementation of a ReAct (Reasoning + Acting) agent capable of self-critique. If information is missing, the agent does not "guess"; it requests human validation (HITL).

Vector Database (Qdrant): Management of hybrid collections (1500+ vectors) with semantic search by cosine similarity on 1536D vectors.

Advanced Embedding Strategy: Use of SLM models (Qwen 2.5) to generate "Enhanced Texts" (summaries and keywords) before indexing, drastically increasing retrieval accuracy.

Results & Added Value

Productivity Gain: Drastic reduction in initial drafting time.

Brand Consistency: Automatic alignment with responses previously validated by business experts.

Full Traceability: Each response is sourced (origin from technical documentation or history).