Hi, I'm Emad.
I build intelligent systems.

Contributed to building CLEAR Investigate, Thomson Reuters' first agentic AI product, live in production for customers. Designed the agent experimentation architecture using PydanticAI for rapid prototyping of multi-step workflows across entity search, report retrieval, web search, and an internal knowledge graph. Implemented LLM caching (~25% cost reduction), automated LLM-as-judge evaluation pipelines, and SME annotation workflows for ground-truth benchmarking.

Built the AI-powered summary panel for CLEAR Business entity reports, live in production processing hundreds of reports daily. Designed a selective XML parsing engine and multi-prompt LLM architecture with separate calls for business overview, risk analysis, and social media discovery — with Bing Search integration and a verification LLM call to filter false positives. Implemented citation linking to source locations and ran SME annotation and evaluation rounds before launch.

Designed and built the data engineering infrastructure for an entity resolution system operating across ~800M entities and billions of documents at 700 idents/second. Architected a PII-isolated AWS environment and a unified versioned schema reconciling two incompatible annotation sources spanning hundreds of thousands of labeled records. Built Spark and Python pipelines for multi-format merging, conflict signal surfacing, and model experimentation, later automated into SageMaker Pipelines.

Shipped a semantic search improvement for Checkpoint, Thomson Reuters' tax research platform, improving access to conceptual documents by 95%. Built a query intent classifier using sentence embeddings to dynamically promote higher-level content for general queries, delivered to production within a ~100ms latency budget.

Fine-tuned T5 transformer for automated question generation, cutting manual data curation effort by 40%. Experimented with model input representations and evaluated using BLEURT as part of a university collaborative research project.

Trained domain-specific Word2Vec embeddings and LSTM-based NLU models with interpretability testing to improve language understanding accuracy and model transparency.

Built a domain-specific question answering dialogue system using syntactic and semantic document analysis and ontology generation, as part of a collaborative industry research project.