Ihjass Thasbekha

• Led design and development of Loan Document Processing (LDP) – Enterprise Generative AI Platform, architecting GPT-4, Gemini, BERT, OCR, and NLP pipelines to automate classification, extraction, and validation across loan applications, bank statements, KYC, and income documents — scaling to 130 documents/week at 95% extraction accuracy.
• Designed prompt engineering and AI-assisted validation workflows that cut downstream data-correction requests by 80% and reduced manual review time from 2 weeks to a single day, with 98% accuracy.
• Identified validation leakage in the admin/processor workflow and built an automated MongoDB-backed validation module, slashing manual verification effort from ~1 day per document to 5–10 minutes.
• Led development of Intelligent Document Analysis (IDA) – Enterprise Generative AI Platform, a Generative AI platform combining LangGraph-based NL-to-SQL analytics and Qdrant-backed hybrid RAG, enabling business users to query Snowflake financial data (loans, deals, tranches) and securitization documents (prospectuses, servicing reports) in natural language.
• Built the end-to-end NL→SQL pipeline (semantic model selection → SQL generation → execution → visualization), enforcing SQL-injection guardrails that restricted all LLM-generated queries to read-only SELECT statements.
• Oversaw the RAG system's document intelligence capabilities — multi-document Q&A, clause extraction, cross-deal comparison, contract risk identification, and OCR-based table extraction — built on Azure OpenAI (GPT-4.1) with Vault-based JWT authentication for secure Snowflake access. Designed interactive Plotly dashboards to visualize AI-generated query results, cutting time-to-insight for business users from 2 hours to 10 minutes.
• Set up CI/CD (GitHub Actions → Docker → self-hosted deployment), enabling reliable, repeatable production releases.
• Directed cross-functional delivery of production-ready AI features and mentored 4 junior engineers, sustaining on-time sp

Skills: Generative AI & LLMs: Azure OpenAI, GPT-4/4.1, Gemini, RAG, Prompt Engineering, LangChain, LangGraph, LlamaIndex, Haystack, AI Agents ML / NLP: BERT, DistilBERT, RCNN, DETR, NER, OCR

• Developed an enterprise AI solution using Gemini and LlamaIndex for structured extraction from unstructured legal filings on a Government of India engagement; introduced GPT-based models into the pipeline, lifting accuracy to 96% and shaping the team's model-adoption roadmap.
• Engineered automated preprocessing (deskewing, page unwrapping) that improved document readability ahead of model inference, adding a further 4% to extraction accuracy.

• Built a document intelligence pipeline extracting structured data from government-issued KYC documents (Aadhaar, passport, driving license) for a major banking client's onboarding workflow, under a strict sub-3-second response requirement.
• Benchmarked extraction architectures from CNN and RCNN through BERT and DistilBERT, then optimized retrieval with annotation-driven preprocessing and the Haystack framework — cutting end-to-end response time to ~1 second.
• Improved OCR-based handling of poor-quality scans, lifting extraction accuracy by 16%.